Upload 9 files

audio.py

@@ -0,0 +1,119 @@
+"""We use the PyAV library to decode the audio: https://github.com/PyAV-Org/PyAV
+
+The advantage of PyAV is that it bundles the FFmpeg libraries so there is no additional
+system dependencies. FFmpeg does not need to be installed on the system.
+
+However, the API is quite low-level so we need to manipulate audio frames directly.
+"""
+
+import gc
+import io
+import itertools
+
+from typing import BinaryIO, Union
+
+import av
+import numpy as np
+
+
+def decode_audio(
+    input_file: Union[str, BinaryIO],
+    sampling_rate: int = 16000,
+    split_stereo: bool = False,
+):
+    """Decodes the audio.
+
+    Args:
+      input_file: Path to the input file or a file-like object.
+      sampling_rate: Resample the audio to this sample rate.
+      split_stereo: Return separate left and right channels.
+
+    Returns:
+      A float32 Numpy array.
+
+      If `split_stereo` is enabled, the function returns a 2-tuple with the
+      separated left and right channels.
+    """
+    resampler = av.audio.resampler.AudioResampler(
+        format="s16",
+        layout="mono" if not split_stereo else "stereo",
+        rate=sampling_rate,
+    )
+
+    raw_buffer = io.BytesIO()
+    dtype = None
+
+    with av.open(input_file, mode="r", metadata_errors="ignore") as container:
+        frames = container.decode(audio=0)
+        frames = _ignore_invalid_frames(frames)
+        frames = _group_frames(frames, 500000)
+        frames = _resample_frames(frames, resampler)
+
+        for frame in frames:
+            array = frame.to_ndarray()
+            dtype = array.dtype
+            raw_buffer.write(array)
+
+    # It appears that some objects related to the resampler are not freed
+    # unless the garbage collector is manually run.
+    del resampler
+    gc.collect()
+
+    audio = np.frombuffer(raw_buffer.getbuffer(), dtype=dtype)
+
+    # Convert s16 back to f32.
+    audio = audio.astype(np.float32) / 32768.0
+
+    if split_stereo:
+        left_channel = audio[0::2]
+        right_channel = audio[1::2]
+        return left_channel, right_channel
+
+    return audio
+
+
+def _ignore_invalid_frames(frames):
+    iterator = iter(frames)
+
+    while True:
+        try:
+            yield next(iterator)
+        except StopIteration:
+            break
+        except av.error.InvalidDataError:
+            continue
+
+
+def _group_frames(frames, num_samples=None):
+    fifo = av.audio.fifo.AudioFifo()
+
+    for frame in frames:
+        frame.pts = None  # Ignore timestamp check.
+        fifo.write(frame)
+
+        if num_samples is not None and fifo.samples >= num_samples:
+            yield fifo.read()
+
+    if fifo.samples > 0:
+        yield fifo.read()
+
+
+def _resample_frames(frames, resampler):
+    # Add None to flush the resampler.
+    for frame in itertools.chain(frames, [None]):
+        yield from resampler.resample(frame)
+
+
+def pad_or_trim(array, length: int, *, axis: int = -1):
+    """
+    Pad or trim the audio array to N_SAMPLES, as expected by the encoder.
+    """
+    if array.shape[axis] > length:
+        array = array.take(indices=range(length), axis=axis)
+
+    if array.shape[axis] < length:
+        pad_widths = [(0, 0)] * array.ndim
+        pad_widths[axis] = (0, length - array.shape[axis])
+        array = np.pad(array, pad_widths)
+
+    return array

feature_extractor.py

@@ -0,0 +1,170 @@
+import numpy as np
+
+
+# Adapted from https://github.com/huggingface/transformers/blob/main/src/transformers/models/whisper/feature_extraction_whisper.py  # noqa: E501
+class FeatureExtractor:
+    def __init__(
+        self,
+        feature_size=80,
+        sampling_rate=16000,
+        hop_length=160,
+        chunk_length=30,
+        n_fft=400,
+    ):
+        self.n_fft = n_fft
+        self.hop_length = hop_length
+        self.chunk_length = chunk_length
+        self.n_samples = chunk_length * sampling_rate
+        self.nb_max_frames = self.n_samples // hop_length
+        self.time_per_frame = hop_length / sampling_rate
+        self.sampling_rate = sampling_rate
+        self.mel_filters = self.get_mel_filters(
+            sampling_rate, n_fft, n_mels=feature_size
+        )
+
+    def get_mel_filters(self, sr, n_fft, n_mels=128, dtype=np.float32):
+        # Initialize the weights
+        n_mels = int(n_mels)
+        weights = np.zeros((n_mels, int(1 + n_fft // 2)), dtype=dtype)
+
+        # Center freqs of each FFT bin
+        fftfreqs = np.fft.rfftfreq(n=n_fft, d=1.0 / sr)
+
+        # 'Center freqs' of mel bands - uniformly spaced between limits
+        min_mel = 0.0
+        max_mel = 45.245640471924965
+
+        mels = np.linspace(min_mel, max_mel, n_mels + 2)
+
+        mels = np.asanyarray(mels)
+
+        # Fill in the linear scale
+        f_min = 0.0
+        f_sp = 200.0 / 3
+        freqs = f_min + f_sp * mels
+
+        # And now the nonlinear scale
+        min_log_hz = 1000.0  # beginning of log region (Hz)
+        min_log_mel = (min_log_hz - f_min) / f_sp  # same (Mels)
+        logstep = np.log(6.4) / 27.0  # step size for log region
+
+        # If we have vector data, vectorize
+        log_t = mels >= min_log_mel
+        freqs[log_t] = min_log_hz * np.exp(logstep * (mels[log_t] - min_log_mel))
+
+        mel_f = freqs
+
+        fdiff = np.diff(mel_f)
+        ramps = np.subtract.outer(mel_f, fftfreqs)
+
+        for i in range(n_mels):
+            # lower and upper slopes for all bins
+            lower = -ramps[i] / fdiff[i]
+            upper = ramps[i + 2] / fdiff[i + 1]
+
+            # .. then intersect them with each other and zero
+            weights[i] = np.maximum(0, np.minimum(lower, upper))
+
+        # Slaney-style mel is scaled to be approx constant energy per channel
+        enorm = 2.0 / (mel_f[2 : n_mels + 2] - mel_f[:n_mels])
+        weights *= enorm[:, np.newaxis]
+
+        return weights
+
+    def fram_wave(self, waveform, center=True):
+        """
+        Transform a raw waveform into a list of smaller waveforms.
+        The window length defines how much of the signal is
+        contain in each frame (smalle waveform), while the hope length defines the step
+        between the beginning of each new frame.
+        Centering is done by reflecting the waveform which is first centered around
+        `frame_idx * hop_length`.
+        """
+        frames = []
+        for i in range(0, waveform.shape[0] + 1, self.hop_length):
+            half_window = (self.n_fft - 1) // 2 + 1
+            if center:
+                start = i - half_window if i > half_window else 0
+                end = (
+                    i + half_window
+                    if i < waveform.shape[0] - half_window
+                    else waveform.shape[0]
+                )
+
+                frame = waveform[start:end]
+
+                if start == 0:
+                    padd_width = (-i + half_window, 0)
+                    frame = np.pad(frame, pad_width=padd_width, mode="reflect")
+
+                elif end == waveform.shape[0]:
+                    padd_width = (0, (i - waveform.shape[0] + half_window))
+                    frame = np.pad(frame, pad_width=padd_width, mode="reflect")
+
+            else:
+                frame = waveform[i : i + self.n_fft]
+                frame_width = frame.shape[0]
+                if frame_width < waveform.shape[0]:
+                    frame = np.lib.pad(
+                        frame,
+                        pad_width=(0, self.n_fft - frame_width),
+                        mode="constant",
+                        constant_values=0,
+                    )
+
+            frames.append(frame)
+        return np.stack(frames, 0)
+
+    def stft(self, frames, window):
+        """
+        Calculates the complex Short-Time Fourier Transform (STFT) of the given framed signal.
+        Should give the same results as `torch.stft`.
+        """
+        frame_size = frames.shape[1]
+        fft_size = self.n_fft
+
+        if fft_size is None:
+            fft_size = frame_size
+
+        if fft_size < frame_size:
+            raise ValueError("FFT size must greater or equal the frame size")
+        # number of FFT bins to store
+        num_fft_bins = (fft_size >> 1) + 1
+
+        data = np.empty((len(frames), num_fft_bins), dtype=np.complex64)
+        fft_signal = np.zeros(fft_size)
+
+        for f, frame in enumerate(frames):
+            if window is not None:
+                np.multiply(frame, window, out=fft_signal[:frame_size])
+            else:
+                fft_signal[:frame_size] = frame
+            data[f] = np.fft.fft(fft_signal, axis=0)[:num_fft_bins]
+        return data.T
+
+    def __call__(self, waveform, padding=True, chunk_length=None):
+        """
+        Compute the log-Mel spectrogram of the provided audio, gives similar results
+        whisper's original torch implementation with 1e-5 tolerance.
+        """
+        if chunk_length is not None:
+            self.n_samples = chunk_length * self.sampling_rate
+            self.nb_max_frames = self.n_samples // self.hop_length
+
+        if padding:
+            waveform = np.pad(waveform, [(0, self.n_samples)])
+
+        window = np.hanning(self.n_fft + 1)[:-1]
+
+        frames = self.fram_wave(waveform)
+        stft = self.stft(frames, window=window)
+        magnitudes = np.abs(stft[:, :-1]) ** 2
+
+        filters = self.mel_filters
+        mel_spec = filters @ magnitudes
+
+        log_spec = np.log10(np.clip(mel_spec, a_min=1e-10, a_max=None))
+        log_spec = np.maximum(log_spec, log_spec.max() - 8.0)
+        log_spec = (log_spec + 4.0) / 4.0
+
+        return log_spec

silero_vad.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:591f853590d11ddde2f2a54f9e7ccecb2533a8af7716330e8adfa6f3849787a9
+size 1807524

tokenizer.py

@@ -0,0 +1,278 @@
+import string
+
+from functools import cached_property
+from typing import List, Optional, Tuple
+
+import tokenizers
+
+
+class Tokenizer:
+    """Simple wrapper around a tokenizers.Tokenizer."""
+
+    def __init__(
+        self,
+        tokenizer: tokenizers.Tokenizer,
+        multilingual: bool,
+        task: Optional[str] = None,
+        language: Optional[str] = None,
+    ):
+        self.tokenizer = tokenizer
+
+        if multilingual:
+            if task not in _TASKS:
+                raise ValueError(
+                    "'%s' is not a valid task (accepted tasks: %s)"
+                    % (task, ", ".join(_TASKS))
+                )
+
+            if language not in _LANGUAGE_CODES:
+                raise ValueError(
+                    "'%s' is not a valid language code (accepted language codes: %s)"
+                    % (language, ", ".join(_LANGUAGE_CODES))
+                )
+
+            self.task = self.tokenizer.token_to_id("<|%s|>" % task)
+            self.language = self.tokenizer.token_to_id("<|%s|>" % language)
+            self.language_code = language
+        else:
+            self.task = None
+            self.language = None
+            self.language_code = "en"
+
+    @cached_property
+    def transcribe(self) -> int:
+        return self.tokenizer.token_to_id("<|transcribe|>")
+
+    @cached_property
+    def translate(self) -> int:
+        return self.tokenizer.token_to_id("<|translate|>")
+
+    @cached_property
+    def sot(self) -> int:
+        return self.tokenizer.token_to_id("<|startoftranscript|>")
+
+    @cached_property
+    def sot_lm(self) -> int:
+        return self.tokenizer.token_to_id("<|startoflm|>")
+
+    @cached_property
+    def sot_prev(self) -> int:
+        return self.tokenizer.token_to_id("<|startofprev|>")
+
+    @cached_property
+    def eot(self) -> int:
+        return self.tokenizer.token_to_id("<|endoftext|>")
+
+    @cached_property
+    def no_timestamps(self) -> int:
+        return self.tokenizer.token_to_id("<|notimestamps|>")
+
+    @property
+    def timestamp_begin(self) -> int:
+        return self.no_timestamps + 1
+
+    @property
+    def sot_sequence(self) -> List[int]:
+        sequence = [self.sot]
+
+        if self.language is not None:
+            sequence.append(self.language)
+
+        if self.task is not None:
+            sequence.append(self.task)
+
+        return sequence
+
+    def encode(self, text: str) -> List[int]:
+        return self.tokenizer.encode(text, add_special_tokens=False).ids
+
+    def decode(self, tokens: List[int]) -> str:
+        text_tokens = [token for token in tokens if token < self.eot]
+        return self.tokenizer.decode(text_tokens)
+
+    def decode_with_timestamps(self, tokens: List[int]) -> str:
+        outputs = [[]]
+
+        for token in tokens:
+            if token >= self.timestamp_begin:
+                timestamp = f"<|{(token - self.timestamp_begin) * 0.02:.2f}|>"
+                outputs.append(timestamp)
+                outputs.append([])
+            else:
+                outputs[-1].append(token)
+
+        return "".join(
+            [s if isinstance(s, str) else self.tokenizer.decode(s) for s in outputs]
+        )
+
+    def split_to_word_tokens(
+        self, tokens: List[int]
+    ) -> Tuple[List[str], List[List[int]]]:
+        if self.language_code in {"zh", "ja", "th", "lo", "my", "yue"}:
+            # These languages don't typically use spaces, so it is difficult to split words
+            # without morpheme analysis. Here, we instead split words at any
+            # position where the tokens are decoded as valid unicode points
+            return self.split_tokens_on_unicode(tokens)
+
+        return self.split_tokens_on_spaces(tokens)
+
+    def split_tokens_on_unicode(
+        self, tokens: List[int]
+    ) -> Tuple[List[str], List[List[int]]]:
+        decoded_full = self.decode_with_timestamps(tokens)
+        replacement_char = "\ufffd"
+
+        words = []
+        word_tokens = []
+        current_tokens = []
+        unicode_offset = 0
+
+        for token in tokens:
+            current_tokens.append(token)
+            decoded = self.decode_with_timestamps(current_tokens)
+
+            try:
+                replacement_char_index = decoded.index(replacement_char)
+                replacement_char_index += unicode_offset
+            except ValueError:
+                replacement_char_index = None
+
+            if replacement_char_index is None or (
+                replacement_char_index < len(decoded_full)
+                and decoded_full[replacement_char_index] == replacement_char
+            ):
+                words.append(decoded)
+                word_tokens.append(current_tokens)
+                current_tokens = []
+                unicode_offset += len(decoded)
+
+        return words, word_tokens
+
+    def split_tokens_on_spaces(
+        self, tokens: List[int]
+    ) -> Tuple[List[str], List[List[int]]]:
+        subwords, subword_tokens_list = self.split_tokens_on_unicode(tokens)
+        words = []
+        word_tokens = []
+
+        for subword, subword_tokens in zip(subwords, subword_tokens_list):
+            special = subword_tokens[0] >= self.eot
+            with_space = subword.startswith(" ")
+            punctuation = subword.strip() in string.punctuation
+            if special or with_space or punctuation or len(words) == 0:
+                words.append(subword)
+                word_tokens.append(subword_tokens)
+            else:
+                words[-1] = words[-1] + subword
+                word_tokens[-1].extend(subword_tokens)
+
+        return words, word_tokens
+
+
+_TASKS = (
+    "transcribe",
+    "translate",
+)
+
+_LANGUAGE_CODES = (
+    "af",
+    "am",
+    "ar",
+    "as",
+    "az",
+    "ba",
+    "be",
+    "bg",
+    "bn",
+    "bo",
+    "br",
+    "bs",
+    "ca",
+    "cs",
+    "cy",
+    "da",
+    "de",
+    "el",
+    "en",
+    "es",
+    "et",
+    "eu",
+    "fa",
+    "fi",
+    "fo",
+    "fr",
+    "gl",
+    "gu",
+    "ha",
+    "haw",
+    "he",
+    "hi",
+    "hr",
+    "ht",
+    "hu",
+    "hy",
+    "id",
+    "is",
+    "it",
+    "ja",
+    "jw",
+    "ka",
+    "kk",
+    "km",
+    "kn",
+    "ko",
+    "la",
+    "lb",
+    "ln",
+    "lo",
+    "lt",
+    "lv",
+    "mg",
+    "mi",
+    "mk",
+    "ml",
+    "mn",
+    "mr",
+    "ms",
+    "mt",
+    "my",
+    "ne",
+    "nl",
+    "nn",
+    "no",
+    "oc",
+    "pa",
+    "pl",
+    "ps",
+    "pt",
+    "ro",
+    "ru",
+    "sa",
+    "sd",
+    "si",
+    "sk",
+    "sl",
+    "sn",
+    "so",
+    "sq",
+    "sr",
+    "su",
+    "sv",
+    "sw",
+    "ta",
+    "te",
+    "tg",
+    "th",
+    "tk",
+    "tl",
+    "tr",
+    "tt",
+    "uk",
+    "ur",
+    "uz",
+    "vi",
+    "yi",
+    "yo",
+    "zh",
+    "yue",
+)

transcribe.py

@@ -0,0 +1,1272 @@
+import itertools
+import json
+import logging
+import os
+import zlib
+
+from inspect import signature
+from typing import BinaryIO, Iterable, List, NamedTuple, Optional, Tuple, Union
+
+import ctranslate2
+import numpy as np
+import tokenizers
+
+from faster_whisper.audio import decode_audio, pad_or_trim
+from faster_whisper.feature_extractor import FeatureExtractor
+from faster_whisper.tokenizer import _LANGUAGE_CODES, Tokenizer
+from faster_whisper.utils import download_model, format_timestamp, get_end, get_logger
+from faster_whisper.vad import (
+    SpeechTimestampsMap,
+    VadOptions,
+    collect_chunks,
+    get_speech_timestamps,
+)
+
+
+class Word(NamedTuple):
+    start: float
+    end: float
+    word: str
+    probability: float
+
+
+class Segment(NamedTuple):
+    id: int
+    seek: int
+    start: float
+    end: float
+    text: str
+    tokens: List[int]
+    temperature: float
+    avg_logprob: float
+    compression_ratio: float
+    no_speech_prob: float
+    words: Optional[List[Word]]
+
+
+class TranscriptionOptions(NamedTuple):
+    beam_size: int
+    best_of: int
+    patience: float
+    length_penalty: float
+    repetition_penalty: float
+    no_repeat_ngram_size: int
+    log_prob_threshold: Optional[float]
+    no_speech_threshold: Optional[float]
+    compression_ratio_threshold: Optional[float]
+    condition_on_previous_text: bool
+    prompt_reset_on_temperature: float
+    temperatures: List[float]
+    initial_prompt: Optional[Union[str, Iterable[int]]]
+    prefix: Optional[str]
+    suppress_blank: bool
+    suppress_tokens: Optional[List[int]]
+    without_timestamps: bool
+    max_initial_timestamp: float
+    word_timestamps: bool
+    prepend_punctuations: str
+    append_punctuations: str
+    max_new_tokens: Optional[int]
+    clip_timestamps: Union[str, List[float]]
+    hallucination_silence_threshold: Optional[float]
+    hotwords: Optional[str]
+
+
+class TranscriptionInfo(NamedTuple):
+    language: str
+    language_probability: float
+    duration: float
+    duration_after_vad: float
+    all_language_probs: Optional[List[Tuple[str, float]]]
+    transcription_options: TranscriptionOptions
+    vad_options: VadOptions
+
+
+class WhisperModel:
+    def __init__(
+        self,
+        model_size_or_path: str,
+        device: str = "auto",
+        device_index: Union[int, List[int]] = 0,
+        compute_type: str = "default",
+        cpu_threads: int = 0,
+        num_workers: int = 1,
+        download_root: Optional[str] = None,
+        local_files_only: bool = False,
+        files: dict = None,
+        **model_kwargs,
+    ):
+        """Initializes the Whisper model.
+
+        Args:
+          model_size_or_path: Size of the model to use (tiny, tiny.en, base, base.en,
+            small, small.en, distil-small.en, medium, medium.en, distil-medium.en, large-v1,
+            large-v2, large-v3, large, distil-large-v2 or distil-large-v3), a path to a
+            converted model directory, or a CTranslate2-converted Whisper model ID from the HF Hub.
+            When a size or a model ID is configured, the converted model is downloaded
+            from the Hugging Face Hub.
+          device: Device to use for computation ("cpu", "cuda", "auto").
+          device_index: Device ID to use.
+            The model can also be loaded on multiple GPUs by passing a list of IDs
+            (e.g. [0, 1, 2, 3]). In that case, multiple transcriptions can run in parallel
+            when transcribe() is called from multiple Python threads (see also num_workers).
+          compute_type: Type to use for computation.
+            See https://opennmt.net/CTranslate2/quantization.html.
+          cpu_threads: Number of threads to use when running on CPU (4 by default).
+            A non zero value overrides the OMP_NUM_THREADS environment variable.
+          num_workers: When transcribe() is called from multiple Python threads,
+            having multiple workers enables true parallelism when running the model
+            (concurrent calls to self.model.generate() will run in parallel).
+            This can improve the global throughput at the cost of increased memory usage.
+          download_root: Directory where the models should be saved. If not set, the models
+            are saved in the standard Hugging Face cache directory.
+          local_files_only:  If True, avoid downloading the file and return the path to the
+            local cached file if it exists.
+          files: Load model files from the memory. This argument is a dictionary mapping file names
+            to file contents as file-like or bytes objects. If this is set, model_path acts as an
+            identifier for this model.
+        """
+        self.logger = get_logger()
+
+        tokenizer_bytes, preprocessor_bytes = None, None
+        if files:
+            model_path = model_size_or_path
+            tokenizer_bytes = files.pop("tokenizer.json", None)
+            preprocessor_bytes = files.pop("preprocessor_config.json", None)
+        elif os.path.isdir(model_size_or_path):
+            model_path = model_size_or_path
+        else:
+            model_path = download_model(
+                model_size_or_path,
+                local_files_only=local_files_only,
+                cache_dir=download_root,
+            )
+
+        self.model = ctranslate2.models.Whisper(
+            model_path,
+            device=device,
+            device_index=device_index,
+            compute_type=compute_type,
+            intra_threads=cpu_threads,
+            inter_threads=num_workers,
+            files=files,
+            **model_kwargs,
+        )
+
+        tokenizer_file = os.path.join(model_path, "tokenizer.json")
+        if tokenizer_bytes:
+            self.hf_tokenizer = tokenizers.Tokenizer.from_buffer(tokenizer_bytes)
+        elif os.path.isfile(tokenizer_file):
+            self.hf_tokenizer = tokenizers.Tokenizer.from_file(tokenizer_file)
+        else:
+            self.hf_tokenizer = tokenizers.Tokenizer.from_pretrained(
+                "openai/whisper-tiny" + ("" if self.model.is_multilingual else ".en")
+            )
+        self.feat_kwargs = self._get_feature_kwargs(model_path, preprocessor_bytes)
+        self.feature_extractor = FeatureExtractor(**self.feat_kwargs)
+        self.num_samples_per_token = self.feature_extractor.hop_length * 2
+        self.frames_per_second = (
+            self.feature_extractor.sampling_rate // self.feature_extractor.hop_length
+        )
+        self.tokens_per_second = (
+            self.feature_extractor.sampling_rate // self.num_samples_per_token
+        )
+        self.input_stride = 2
+        self.time_precision = 0.02
+        self.max_length = 448
+
+    @property
+    def supported_languages(self) -> List[str]:
+        """The languages supported by the model."""
+        return list(_LANGUAGE_CODES) if self.model.is_multilingual else ["en"]
+
+    def _get_feature_kwargs(self, model_path, preprocessor_bytes=None) -> dict:
+        config = {}
+        try:
+            config_path = os.path.join(model_path, "preprocessor_config.json")
+            if preprocessor_bytes:
+                config = json.loads(preprocessor_bytes)
+            elif os.path.isfile(config_path):
+                with open(config_path, "r", encoding="utf-8") as file:
+                    config = json.load(file)
+            else:
+                return config
+            valid_keys = signature(FeatureExtractor.__init__).parameters.keys()
+            return {k: v for k, v in config.items() if k in valid_keys}
+        except json.JSONDecodeError as e:
+            self.logger.warning("Could not load preprocessor config: %s", e)
+
+        return config
+
+    def transcribe(
+        self,
+        audio: Union[str, BinaryIO, np.ndarray],
+        language: Optional[str] = None,
+        task: str = "transcribe",
+        beam_size: int = 5,
+        best_of: int = 5,
+        patience: float = 1,
+        length_penalty: float = 1,
+        repetition_penalty: float = 1,
+        no_repeat_ngram_size: int = 0,
+        temperature: Union[float, List[float], Tuple[float, ...]] = [
+            0.0,
+            0.2,
+            0.4,
+            0.6,
+            0.8,
+            1.0,
+        ],
+        compression_ratio_threshold: Optional[float] = 2.4,
+        log_prob_threshold: Optional[float] = -1.0,
+        no_speech_threshold: Optional[float] = 0.6,
+        condition_on_previous_text: bool = True,
+        prompt_reset_on_temperature: float = 0.5,
+        initial_prompt: Optional[Union[str, Iterable[int]]] = None,
+        prefix: Optional[str] = None,
+        suppress_blank: bool = True,
+        suppress_tokens: Optional[List[int]] = [-1],
+        without_timestamps: bool = False,
+        max_initial_timestamp: float = 1.0,
+        word_timestamps: bool = False,
+        prepend_punctuations: str = "\"'“¿([{-",
+        append_punctuations: str = "\"'.。,，!！?？:：”)]}、",
+        vad_filter: bool = False,
+        vad_parameters: Optional[Union[dict, VadOptions]] = None,
+        max_new_tokens: Optional[int] = None,
+        chunk_length: Optional[int] = None,
+        clip_timestamps: Union[str, List[float]] = "0",
+        hallucination_silence_threshold: Optional[float] = None,
+        hotwords: Optional[str] = None,
+        language_detection_threshold: Optional[float] = None,
+        language_detection_segments: int = 1,
+    ) -> Tuple[Iterable[Segment], TranscriptionInfo]:
+        """Transcribes an input file.
+
+        Arguments:
+          audio: Path to the input file (or a file-like object), or the audio waveform.
+          language: The language spoken in the audio. It should be a language code such
+            as "en" or "fr". If not set, the language will be detected in the first 30 seconds
+            of audio.
+          task: Task to execute (transcribe or translate).
+          beam_size: Beam size to use for decoding.
+          best_of: Number of candidates when sampling with non-zero temperature.
+          patience: Beam search patience factor.
+          length_penalty: Exponential length penalty constant.
+          repetition_penalty: Penalty applied to the score of previously generated tokens
+            (set > 1 to penalize).
+          no_repeat_ngram_size: Prevent repetitions of ngrams with this size (set 0 to disable).
+          temperature: Temperature for sampling. It can be a tuple of temperatures,
+            which will be successively used upon failures according to either
+            `compression_ratio_threshold` or `log_prob_threshold`.
+          compression_ratio_threshold: If the gzip compression ratio is above this value,
+            treat as failed.
+          log_prob_threshold: If the average log probability over sampled tokens is
+            below this value, treat as failed.
+          no_speech_threshold: If the no_speech probability is higher than this value AND
+            the average log probability over sampled tokens is below `log_prob_threshold`,
+            consider the segment as silent.
+          condition_on_previous_text: If True, the previous output of the model is provided
+            as a prompt for the next window; disabling may make the text inconsistent across
+            windows, but the model becomes less prone to getting stuck in a failure loop,
+            such as repetition looping or timestamps going out of sync.
+          prompt_reset_on_temperature: Resets prompt if temperature is above this value.
+            Arg has effect only if condition_on_previous_text is True.
+          initial_prompt: Optional text string or iterable of token ids to provide as a
+            prompt for the first window.
+          prefix: Optional text to provide as a prefix for the first window.
+          suppress_blank: Suppress blank outputs at the beginning of the sampling.
+          suppress_tokens: List of token IDs to suppress. -1 will suppress a default set
+            of symbols as defined in the model config.json file.
+          without_timestamps: Only sample text tokens.
+          max_initial_timestamp: The initial timestamp cannot be later than this.
+          word_timestamps: Extract word-level timestamps using the cross-attention pattern
+            and dynamic time warping, and include the timestamps for each word in each segment.
+          prepend_punctuations: If word_timestamps is True, merge these punctuation symbols
+            with the next word
+          append_punctuations: If word_timestamps is True, merge these punctuation symbols
+            with the previous word
+          vad_filter: Enable the voice activity detection (VAD) to filter out parts of the audio
+            without speech. This step is using the Silero VAD model
+            https://github.com/snakers4/silero-vad.
+          vad_parameters: Dictionary of Silero VAD parameters or VadOptions class (see available
+            parameters and default values in the class `VadOptions`).
+          max_new_tokens: Maximum number of new tokens to generate per-chunk. If not set,
+            the maximum will be set by the default max_length.
+          chunk_length: The length of audio segments. If it is not None, it will overwrite the
+            default chunk_length of the FeatureExtractor.
+          clip_timestamps:
+            Comma-separated list start,end,start,end,... timestamps (in seconds) of clips to
+             process. The last end timestamp defaults to the end of the file.
+             vad_filter will be ignored if clip_timestamps is used.
+          hallucination_silence_threshold:
+            When word_timestamps is True, skip silent periods longer than this threshold
+             (in seconds) when a possible hallucination is detected
+          hotwords:
+            Hotwords/hint phrases to provide the model with. Has no effect if prefix is not None.
+          language_detection_threshold: If the maximum probability of the language tokens is higher
+           than this value, the language is detected.
+          language_detection_segments: Number of segments to consider for the language detection.
+        Returns:
+          A tuple with:
+
+            - a generator over transcribed segments
+            - an instance of TranscriptionInfo
+        """
+        sampling_rate = self.feature_extractor.sampling_rate
+
+        if not isinstance(audio, np.ndarray):
+            audio = decode_audio(audio, sampling_rate=sampling_rate)
+
+        duration = audio.shape[0] / sampling_rate
+        duration_after_vad = duration
+
+        self.logger.info(
+            "Processing audio with duration %s", format_timestamp(duration)
+        )
+
+        if vad_filter and clip_timestamps == "0":
+            if vad_parameters is None:
+                vad_parameters = VadOptions()
+            elif isinstance(vad_parameters, dict):
+                vad_parameters = VadOptions(**vad_parameters)
+            speech_chunks = get_speech_timestamps(audio, vad_parameters)
+            audio = collect_chunks(audio, speech_chunks)
+            duration_after_vad = audio.shape[0] / sampling_rate
+
+            self.logger.info(
+                "VAD filter removed %s of audio",
+                format_timestamp(duration - duration_after_vad),
+            )
+
+            if self.logger.isEnabledFor(logging.DEBUG):
+                self.logger.debug(
+                    "VAD filter kept the following audio segments: %s",
+                    ", ".join(
+                        "[%s -> %s]"
+                        % (
+                            format_timestamp(chunk["start"] / sampling_rate),
+                            format_timestamp(chunk["end"] / sampling_rate),
+                        )
+                        for chunk in speech_chunks
+                    ),
+                )
+
+        else:
+            speech_chunks = None
+
+        features = self.feature_extractor(audio, chunk_length=chunk_length)
+
+        encoder_output = None
+        all_language_probs = None
+
+        if language is None:
+            if not self.model.is_multilingual:
+                language = "en"
+                language_probability = 1
+            else:
+                if (
+                    language_detection_segments is None
+                    or language_detection_segments < 1
+                ):
+                    language_detection_segments = 1
+                seek = 0
+                detected_language_info = {}
+                content_frames = (
+                    features.shape[-1] - self.feature_extractor.nb_max_frames
+                )
+                while (
+                    seek <= content_frames
+                    and seek
+                    < self.feature_extractor.nb_max_frames * language_detection_segments
+                ):
+                    segment = features[
+                        :, seek : seek + self.feature_extractor.nb_max_frames
+                    ]
+                    encoder_output = self.encode(segment)
+                    # results is a list of tuple[str, float] with language names and
+                    # probabilities.
+                    results = self.model.detect_language(encoder_output)[0]
+                    # Parse language names to strip out markers
+                    all_language_probs = [
+                        (token[2:-2], prob) for (token, prob) in results
+                    ]
+                    # Get top language token and probability
+                    language, language_probability = all_language_probs[0]
+                    if (
+                        language_detection_threshold is None
+                        or language_probability > language_detection_threshold
+                    ):
+                        break
+                    detected_language_info.setdefault(language, []).append(
+                        language_probability
+                    )
+                    seek += segment.shape[-1]
+                else:
+                    # If no language detected for all segments, the majority vote of the highest
+                    # projected languages for all segments is used to determine the language.
+                    language = max(
+                        detected_language_info,
+                        key=lambda lang: len(detected_language_info[lang]),
+                    )
+                    language_probability = max(detected_language_info[language])
+
+                self.logger.info(
+                    "Detected language '%s' with probability %.2f",
+                    language,
+                    language_probability,
+                )
+        else:
+            if not self.model.is_multilingual and language != "en":
+                self.logger.warning(
+                    "The current model is English-only but the language parameter is set to '%s'; "
+                    "using 'en' instead." % language
+                )
+                language = "en"
+
+            language_probability = 1
+
+        tokenizer = Tokenizer(
+            self.hf_tokenizer,
+            self.model.is_multilingual,
+            task=task,
+            language=language,
+        )
+
+        options = TranscriptionOptions(
+            beam_size=beam_size,
+            best_of=best_of,
+            patience=patience,
+            length_penalty=length_penalty,
+            repetition_penalty=repetition_penalty,
+            no_repeat_ngram_size=no_repeat_ngram_size,
+            log_prob_threshold=log_prob_threshold,
+            no_speech_threshold=no_speech_threshold,
+            compression_ratio_threshold=compression_ratio_threshold,
+            condition_on_previous_text=condition_on_previous_text,
+            prompt_reset_on_temperature=prompt_reset_on_temperature,
+            temperatures=(
+                temperature if isinstance(temperature, (list, tuple)) else [temperature]
+            ),
+            initial_prompt=initial_prompt,
+            prefix=prefix,
+            suppress_blank=suppress_blank,
+            suppress_tokens=get_suppressed_tokens(tokenizer, suppress_tokens),
+            without_timestamps=without_timestamps,
+            max_initial_timestamp=max_initial_timestamp,
+            word_timestamps=word_timestamps,
+            prepend_punctuations=prepend_punctuations,
+            append_punctuations=append_punctuations,
+            max_new_tokens=max_new_tokens,
+            clip_timestamps=clip_timestamps,
+            hallucination_silence_threshold=hallucination_silence_threshold,
+            hotwords=hotwords,
+        )
+
+        segments = self.generate_segments(features, tokenizer, options, encoder_output)
+
+        if speech_chunks:
+            segments = restore_speech_timestamps(segments, speech_chunks, sampling_rate)
+
+        info = TranscriptionInfo(
+            language=language,
+            language_probability=language_probability,
+            duration=duration,
+            duration_after_vad=duration_after_vad,
+            transcription_options=options,
+            vad_options=vad_parameters,
+            all_language_probs=all_language_probs,
+        )
+
+        return segments, info
+
+    def generate_segments(
+        self,
+        features: np.ndarray,
+        tokenizer: Tokenizer,
+        options: TranscriptionOptions,
+        encoder_output: Optional[ctranslate2.StorageView] = None,
+    ) -> Iterable[Segment]:
+        content_frames = features.shape[-1] - self.feature_extractor.nb_max_frames
+        content_duration = float(content_frames * self.feature_extractor.time_per_frame)
+
+        if isinstance(options.clip_timestamps, str):
+            options = options._replace(
+                clip_timestamps=[
+                    float(ts)
+                    for ts in (
+                        options.clip_timestamps.split(",")
+                        if options.clip_timestamps
+                        else []
+                    )
+                ]
+            )
+        seek_points: List[int] = [
+            round(ts * self.frames_per_second) for ts in options.clip_timestamps
+        ]
+        if len(seek_points) == 0:
+            seek_points.append(0)
+        if len(seek_points) % 2 == 1:
+            seek_points.append(content_frames)
+        seek_clips: List[Tuple[int, int]] = list(
+            zip(seek_points[::2], seek_points[1::2])
+        )
+
+        punctuation = "\"'“¿([{-\"'.。,，!！?？:：”)]}、"
+
+        idx = 0
+        clip_idx = 0
+        seek = seek_clips[clip_idx][0]
+        all_tokens = []
+        prompt_reset_since = 0
+
+        if options.initial_prompt is not None:
+            if isinstance(options.initial_prompt, str):
+                initial_prompt = " " + options.initial_prompt.strip()
+                initial_prompt_tokens = tokenizer.encode(initial_prompt)
+                all_tokens.extend(initial_prompt_tokens)
+            else:
+                all_tokens.extend(options.initial_prompt)
+
+        last_speech_timestamp = 0.0
+        # NOTE: This loop is obscurely flattened to make the diff readable.
+        # A later commit should turn this into a simpler nested loop.
+        # for seek_clip_start, seek_clip_end in seek_clips:
+        #     while seek < seek_clip_end
+        while clip_idx < len(seek_clips):
+            seek_clip_start, seek_clip_end = seek_clips[clip_idx]
+            if seek_clip_end > content_frames:
+                seek_clip_end = content_frames
+            if seek < seek_clip_start:
+                seek = seek_clip_start
+            if seek >= seek_clip_end:
+                clip_idx += 1
+                if clip_idx < len(seek_clips):
+                    seek = seek_clips[clip_idx][0]
+                continue
+            time_offset = seek * self.feature_extractor.time_per_frame
+            window_end_time = float(
+                (seek + self.feature_extractor.nb_max_frames)
+                * self.feature_extractor.time_per_frame
+            )
+            segment_size = min(
+                self.feature_extractor.nb_max_frames,
+                content_frames - seek,
+                seek_clip_end - seek,
+            )
+            segment = features[:, seek : seek + segment_size]
+            segment_duration = segment_size * self.feature_extractor.time_per_frame
+            segment = pad_or_trim(segment, self.feature_extractor.nb_max_frames)
+
+            if self.logger.isEnabledFor(logging.DEBUG):
+                self.logger.debug(
+                    "Processing segment at %s", format_timestamp(time_offset)
+                )
+
+            previous_tokens = all_tokens[prompt_reset_since:]
+            prompt = self.get_prompt(
+                tokenizer,
+                previous_tokens,
+                without_timestamps=options.without_timestamps,
+                prefix=options.prefix if seek == 0 else None,
+                hotwords=options.hotwords,
+            )
+
+            if seek > 0 or encoder_output is None:
+                encoder_output = self.encode(segment)
+
+            (
+                result,
+                avg_logprob,
+                temperature,
+                compression_ratio,
+            ) = self.generate_with_fallback(encoder_output, prompt, tokenizer, options)
+
+            if options.no_speech_threshold is not None:
+                # no voice activity check
+                should_skip = result.no_speech_prob > options.no_speech_threshold
+
+                if (
+                    options.log_prob_threshold is not None
+                    and avg_logprob > options.log_prob_threshold
+                ):
+                    # don't skip if the logprob is high enough, despite the no_speech_prob
+                    should_skip = False
+
+                if should_skip:
+                    self.logger.debug(
+                        "No speech threshold is met (%f > %f)",
+                        result.no_speech_prob,
+                        options.no_speech_threshold,
+                    )
+
+                    # fast-forward to the next segment boundary
+                    seek += segment_size
+                    continue
+
+            tokens = result.sequences_ids[0]
+
+            previous_seek = seek
+            current_segments = []
+
+            # anomalous words are very long/short/improbable
+            def word_anomaly_score(word: dict) -> float:
+                probability = word.get("probability", 0.0)
+                duration = word["end"] - word["start"]
+                score = 0.0
+                if probability < 0.15:
+                    score += 1.0
+                if duration < 0.133:
+                    score += (0.133 - duration) * 15
+                if duration > 2.0:
+                    score += duration - 2.0
+                return score
+
+            def is_segment_anomaly(segment: Optional[dict]) -> bool:
+                if segment is None or not segment["words"]:
+                    return False
+                words = [w for w in segment["words"] if w["word"] not in punctuation]
+                words = words[:8]
+                score = sum(word_anomaly_score(w) for w in words)
+                return score >= 3 or score + 0.01 >= len(words)
+
+            def next_words_segment(segments: List[dict]) -> Optional[dict]:
+                return next((s for s in segments if s["words"]), None)
+
+            single_timestamp_ending = (
+                len(tokens) >= 2
+                and tokens[-2] < tokenizer.timestamp_begin <= tokens[-1]
+            )
+
+            consecutive_timestamps = [
+                i
+                for i in range(len(tokens))
+                if i > 0
+                and tokens[i] >= tokenizer.timestamp_begin
+                and tokens[i - 1] >= tokenizer.timestamp_begin
+            ]
+
+            if len(consecutive_timestamps) > 0:
+                slices = list(consecutive_timestamps)
+                if single_timestamp_ending:
+                    slices.append(len(tokens))
+
+                last_slice = 0
+                for current_slice in slices:
+                    sliced_tokens = tokens[last_slice:current_slice]
+                    start_timestamp_position = (
+                        sliced_tokens[0] - tokenizer.timestamp_begin
+                    )
+                    end_timestamp_position = (
+                        sliced_tokens[-1] - tokenizer.timestamp_begin
+                    )
+                    start_time = (
+                        time_offset + start_timestamp_position * self.time_precision
+                    )
+                    end_time = (
+                        time_offset + end_timestamp_position * self.time_precision
+                    )
+
+                    current_segments.append(
+                        dict(
+                            seek=seek,
+                            start=start_time,
+                            end=end_time,
+                            tokens=sliced_tokens,
+                        )
+                    )
+                    last_slice = current_slice
+
+                if single_timestamp_ending:
+                    # single timestamp at the end means no speech after the last timestamp.
+                    seek += segment_size
+                else:
+                    # otherwise, ignore the unfinished segment and seek to the last timestamp
+                    last_timestamp_position = (
+                        tokens[last_slice - 1] - tokenizer.timestamp_begin
+                    )
+                    seek += last_timestamp_position * self.input_stride
+
+            else:
+                duration = segment_duration
+                timestamps = [
+                    token for token in tokens if token >= tokenizer.timestamp_begin
+                ]
+                if len(timestamps) > 0 and timestamps[-1] != tokenizer.timestamp_begin:
+                    last_timestamp_position = timestamps[-1] - tokenizer.timestamp_begin
+                    duration = last_timestamp_position * self.time_precision
+
+                current_segments.append(
+                    dict(
+                        seek=seek,
+                        start=time_offset,
+                        end=time_offset + duration,
+                        tokens=tokens,
+                    )
+                )
+
+                seek += segment_size
+
+            if options.word_timestamps:
+                self.add_word_timestamps(
+                    current_segments,
+                    tokenizer,
+                    encoder_output,
+                    segment_size,
+                    options.prepend_punctuations,
+                    options.append_punctuations,
+                    last_speech_timestamp=last_speech_timestamp,
+                )
+
+                if not single_timestamp_ending:
+                    last_word_end = get_end(current_segments)
+                    if last_word_end is not None and last_word_end > time_offset:
+                        seek = round(last_word_end * self.frames_per_second)
+
+                # skip silence before possible hallucinations
+                if options.hallucination_silence_threshold is not None:
+                    threshold = options.hallucination_silence_threshold
+
+                    # if first segment might be a hallucination, skip leading silence
+                    first_segment = next_words_segment(current_segments)
+                    if first_segment is not None and is_segment_anomaly(first_segment):
+                        gap = first_segment["start"] - time_offset
+                        if gap > threshold:
+                            seek = previous_seek + round(gap * self.frames_per_second)
+                            continue
+
+                    # skip silence before any possible hallucination that is surrounded
+                    # by silence or more hallucinations
+                    hal_last_end = last_speech_timestamp
+                    for si in range(len(current_segments)):
+                        segment = current_segments[si]
+                        if not segment["words"]:
+                            continue
+                        if is_segment_anomaly(segment):
+                            next_segment = next_words_segment(
+                                current_segments[si + 1 :]
+                            )
+                            if next_segment is not None:
+                                hal_next_start = next_segment["words"][0]["start"]
+                            else:
+                                hal_next_start = time_offset + segment_duration
+                            silence_before = (
+                                segment["start"] - hal_last_end > threshold
+                                or segment["start"] < threshold
+                                or segment["start"] - time_offset < 2.0
+                            )
+                            silence_after = (
+                                hal_next_start - segment["end"] > threshold
+                                or is_segment_anomaly(next_segment)
+                                or window_end_time - segment["end"] < 2.0
+                            )
+                            if silence_before and silence_after:
+                                seek = round(
+                                    max(time_offset + 1, segment["start"])
+                                    * self.frames_per_second
+                                )
+                                if content_duration - segment["end"] < threshold:
+                                    seek = content_frames
+                                current_segments[si:] = []
+                                break
+                        hal_last_end = segment["end"]
+
+                last_word_end = get_end(current_segments)
+                if last_word_end is not None:
+                    last_speech_timestamp = last_word_end
+
+            for segment in current_segments:
+                tokens = segment["tokens"]
+                text = tokenizer.decode(tokens)
+
+                if segment["start"] == segment["end"] or not text.strip():
+                    continue
+
+                all_tokens.extend(tokens)
+                idx += 1
+
+                yield Segment(
+                    id=idx,
+                    seek=seek,
+                    start=segment["start"],
+                    end=segment["end"],
+                    text=text,
+                    tokens=tokens,
+                    temperature=temperature,
+                    avg_logprob=avg_logprob,
+                    compression_ratio=compression_ratio,
+                    no_speech_prob=result.no_speech_prob,
+                    words=(
+                        [Word(**word) for word in segment["words"]]
+                        if options.word_timestamps
+                        else None
+                    ),
+                )
+
+            if (
+                not options.condition_on_previous_text
+                or temperature > options.prompt_reset_on_temperature
+            ):
+                if options.condition_on_previous_text:
+                    self.logger.debug(
+                        "Reset prompt. prompt_reset_on_temperature threshold is met %f > %f",
+                        temperature,
+                        options.prompt_reset_on_temperature,
+                    )
+
+                prompt_reset_since = len(all_tokens)
+
+    def encode(self, features: np.ndarray) -> ctranslate2.StorageView:
+        # When the model is running on multiple GPUs, the encoder output should be moved
+        # to the CPU since we don't know which GPU will handle the next job.
+        to_cpu = self.model.device == "cuda" and len(self.model.device_index) > 1
+
+        features = np.expand_dims(features, 0)
+        features = get_ctranslate2_storage(features)
+
+        return self.model.encode(features, to_cpu=to_cpu)
+
+    def generate_with_fallback(
+        self,
+        encoder_output: ctranslate2.StorageView,
+        prompt: List[int],
+        tokenizer: Tokenizer,
+        options: TranscriptionOptions,
+    ) -> Tuple[ctranslate2.models.WhisperGenerationResult, float, float, float]:
+        decode_result = None
+        all_results = []
+        below_cr_threshold_results = []
+
+        max_initial_timestamp_index = int(
+            round(options.max_initial_timestamp / self.time_precision)
+        )
+        if options.max_new_tokens is not None:
+            max_length = len(prompt) + options.max_new_tokens
+        else:
+            max_length = self.max_length
+
+        if max_length > self.max_length:
+            raise ValueError(
+                f"The length of the prompt is {len(prompt)}, and the `max_new_tokens` "
+                f"{max_length - len(prompt)}. Thus, the combined length of the prompt "
+                f"and `max_new_tokens` is: {max_length}. This exceeds the "
+                f"`max_length` of the Whisper model: {self.max_length}. "
+                "You should either reduce the length of your prompt, or "
+                "reduce the value of `max_new_tokens`, "
+                f"so that their combined length is less that {self.max_length}."
+            )
+
+        for temperature in options.temperatures:
+            if temperature > 0:
+                kwargs = {
+                    "beam_size": 1,
+                    "num_hypotheses": options.best_of,
+                    "sampling_topk": 0,
+                    "sampling_temperature": temperature,
+                }
+            else:
+                kwargs = {
+                    "beam_size": options.beam_size,
+                    "patience": options.patience,
+                }
+
+            result = self.model.generate(
+                encoder_output,
+                [prompt],
+                length_penalty=options.length_penalty,
+                repetition_penalty=options.repetition_penalty,
+                no_repeat_ngram_size=options.no_repeat_ngram_size,
+                max_length=max_length,
+                return_scores=True,
+                return_no_speech_prob=True,
+                suppress_blank=options.suppress_blank,
+                suppress_tokens=options.suppress_tokens,
+                max_initial_timestamp_index=max_initial_timestamp_index,
+                **kwargs,
+            )[0]
+
+            tokens = result.sequences_ids[0]
+
+            # Recover the average log prob from the returned score.
+            seq_len = len(tokens)
+            cum_logprob = result.scores[0] * (seq_len**options.length_penalty)
+            avg_logprob = cum_logprob / (seq_len + 1)
+
+            text = tokenizer.decode(tokens).strip()
+            compression_ratio = get_compression_ratio(text)
+
+            decode_result = (
+                result,
+                avg_logprob,
+                temperature,
+                compression_ratio,
+            )
+            all_results.append(decode_result)
+
+            needs_fallback = False
+
+            if options.compression_ratio_threshold is not None:
+                if compression_ratio > options.compression_ratio_threshold:
+                    needs_fallback = True  # too repetitive
+
+                    self.logger.debug(
+                        "Compression ratio threshold is not met with temperature %.1f (%f > %f)",
+                        temperature,
+                        compression_ratio,
+                        options.compression_ratio_threshold,
+                    )
+                else:
+                    below_cr_threshold_results.append(decode_result)
+
+            if (
+                options.log_prob_threshold is not None
+                and avg_logprob < options.log_prob_threshold
+            ):
+                needs_fallback = True  # average log probability is too low
+
+                self.logger.debug(
+                    "Log probability threshold is not met with temperature %.1f (%f < %f)",
+                    temperature,
+                    avg_logprob,
+                    options.log_prob_threshold,
+                )
+
+            if (
+                options.no_speech_threshold is not None
+                and result.no_speech_prob > options.no_speech_threshold
+                and options.log_prob_threshold is not None
+                and avg_logprob < options.log_prob_threshold
+            ):
+                needs_fallback = False  # silence
+
+            if not needs_fallback:
+                break
+        else:
+            # all failed, select the result with the highest average log probability
+            decode_result = max(
+                below_cr_threshold_results or all_results, key=lambda x: x[1]
+            )
+            # to pass final temperature for prompt_reset_on_temperature
+            decode_result = (
+                decode_result[0],
+                decode_result[1],
+                temperature,
+                decode_result[3],
+            )
+
+        return decode_result
+
+    def get_prompt(
+        self,
+        tokenizer: Tokenizer,
+        previous_tokens: List[int],
+        without_timestamps: bool = False,
+        prefix: Optional[str] = None,
+        hotwords: Optional[str] = None,
+    ) -> List[int]:
+        prompt = []
+
+        if previous_tokens or (hotwords and not prefix):
+            prompt.append(tokenizer.sot_prev)
+            if hotwords and not prefix:
+                hotwords_tokens = tokenizer.encode(" " + hotwords.strip())
+                if len(hotwords_tokens) >= self.max_length // 2:
+                    hotwords_tokens = hotwords_tokens[: self.max_length // 2 - 1]
+                prompt.extend(hotwords_tokens)
+            if previous_tokens:
+                prompt.extend(previous_tokens[-(self.max_length // 2 - 1) :])
+
+        prompt.extend(tokenizer.sot_sequence)
+
+        if without_timestamps:
+            prompt.append(tokenizer.no_timestamps)
+
+        if prefix:
+            prefix_tokens = tokenizer.encode(" " + prefix.strip())
+            if len(prefix_tokens) >= self.max_length // 2:
+                prefix_tokens = prefix_tokens[: self.max_length // 2 - 1]
+            if not without_timestamps:
+                prompt.append(tokenizer.timestamp_begin)
+            prompt.extend(prefix_tokens)
+
+        return prompt
+
+    def add_word_timestamps(
+        self,
+        segments: List[dict],
+        tokenizer: Tokenizer,
+        encoder_output: ctranslate2.StorageView,
+        num_frames: int,
+        prepend_punctuations: str,
+        append_punctuations: str,
+        last_speech_timestamp: float,
+    ) -> None:
+        if len(segments) == 0:
+            return
+
+        text_tokens_per_segment = [
+            [token for token in segment["tokens"] if token < tokenizer.eot]
+            for segment in segments
+        ]
+
+        text_tokens = list(itertools.chain.from_iterable(text_tokens_per_segment))
+        alignment = self.find_alignment(
+            tokenizer, text_tokens, encoder_output, num_frames
+        )
+        word_durations = np.array([word["end"] - word["start"] for word in alignment])
+        word_durations = word_durations[word_durations.nonzero()]
+        median_duration = np.median(word_durations) if len(word_durations) > 0 else 0.0
+        median_duration = min(0.7, float(median_duration))
+        max_duration = median_duration * 2
+
+        # hack: truncate long words at sentence boundaries.
+        # a better segmentation algorithm based on VAD should be able to replace this.
+        if len(word_durations) > 0:
+            sentence_end_marks = ".。!！?？"
+            # ensure words at sentence boundaries
+            # are not longer than twice the median word duration.
+            for i in range(1, len(alignment)):
+                if alignment[i]["end"] - alignment[i]["start"] > max_duration:
+                    if alignment[i]["word"] in sentence_end_marks:
+                        alignment[i]["end"] = alignment[i]["start"] + max_duration
+                    elif alignment[i - 1]["word"] in sentence_end_marks:
+                        alignment[i]["start"] = alignment[i]["end"] - max_duration
+
+        merge_punctuations(alignment, prepend_punctuations, append_punctuations)
+
+        time_offset = (
+            segments[0]["seek"]
+            * self.feature_extractor.hop_length
+            / self.feature_extractor.sampling_rate
+        )
+
+        word_index = 0
+
+        for segment, text_tokens in zip(segments, text_tokens_per_segment):
+            saved_tokens = 0
+            words = []
+
+            while word_index < len(alignment) and saved_tokens < len(text_tokens):
+                timing = alignment[word_index]
+
+                if timing["word"]:
+                    words.append(
+                        dict(
+                            word=timing["word"],
+                            start=round(time_offset + timing["start"], 2),
+                            end=round(time_offset + timing["end"], 2),
+                            probability=timing["probability"],
+                        )
+                    )
+
+                saved_tokens += len(timing["tokens"])
+                word_index += 1
+
+            # hack: truncate long words at segment boundaries.
+            # a better segmentation algorithm based on VAD should be able to replace this.
+            if len(words) > 0:
+                # ensure the first and second word after a pause is not longer than
+                # twice the median word duration.
+                if words[0]["end"] - last_speech_timestamp > median_duration * 4 and (
+                    words[0]["end"] - words[0]["start"] > max_duration
+                    or (
+                        len(words) > 1
+                        and words[1]["end"] - words[0]["start"] > max_duration * 2
+                    )
+                ):
+                    if (
+                        len(words) > 1
+                        and words[1]["end"] - words[1]["start"] > max_duration
+                    ):
+                        boundary = max(
+                            words[1]["end"] / 2, words[1]["end"] - max_duration
+                        )
+                        words[0]["end"] = words[1]["start"] = boundary
+                    words[0]["start"] = max(0, words[0]["end"] - max_duration)
+
+                # prefer the segment-level start timestamp if the first word is too long.
+                if (
+                    segment["start"] < words[0]["end"]
+                    and segment["start"] - 0.5 > words[0]["start"]
+                ):
+                    words[0]["start"] = max(
+                        0, min(words[0]["end"] - median_duration, segment["start"])
+                    )
+                else:
+                    segment["start"] = words[0]["start"]
+
+                # prefer the segment-level end timestamp if the last word is too long.
+                if (
+                    segment["end"] > words[-1]["start"]
+                    and segment["end"] + 0.5 < words[-1]["end"]
+                ):
+                    words[-1]["end"] = max(
+                        words[-1]["start"] + median_duration, segment["end"]
+                    )
+                else:
+                    segment["end"] = words[-1]["end"]
+
+                last_speech_timestamp = segment["end"]
+
+            segment["words"] = words
+
+    def find_alignment(
+        self,
+        tokenizer: Tokenizer,
+        text_tokens: List[int],
+        encoder_output: ctranslate2.StorageView,
+        num_frames: int,
+        median_filter_width: int = 7,
+    ) -> List[dict]:
+        if len(text_tokens) == 0:
+            return []
+
+        result = self.model.align(
+            encoder_output,
+            tokenizer.sot_sequence,
+            [text_tokens],
+            num_frames,
+            median_filter_width=median_filter_width,
+        )[0]
+
+        text_token_probs = result.text_token_probs
+
+        alignments = result.alignments
+        text_indices = np.array([pair[0] for pair in alignments])
+        time_indices = np.array([pair[1] for pair in alignments])
+
+        words, word_tokens = tokenizer.split_to_word_tokens(
+            text_tokens + [tokenizer.eot]
+        )
+        if len(word_tokens) <= 1:
+            # return on eot only
+            # >>> np.pad([], (1, 0))
+            # array([0.])
+            # This results in crashes when we lookup jump_times with float, like
+            # IndexError: arrays used as indices must be of integer (or boolean) type
+            return []
+        word_boundaries = np.pad(np.cumsum([len(t) for t in word_tokens[:-1]]), (1, 0))
+        if len(word_boundaries) <= 1:
+            return []
+
+        jumps = np.pad(np.diff(text_indices), (1, 0), constant_values=1).astype(bool)
+        jump_times = time_indices[jumps] / self.tokens_per_second
+        start_times = jump_times[word_boundaries[:-1]]
+        end_times = jump_times[word_boundaries[1:]]
+        word_probabilities = [
+            np.mean(text_token_probs[i:j])
+            for i, j in zip(word_boundaries[:-1], word_boundaries[1:])
+        ]
+
+        return [
+            dict(
+                word=word, tokens=tokens, start=start, end=end, probability=probability
+            )
+            for word, tokens, start, end, probability in zip(
+                words, word_tokens, start_times, end_times, word_probabilities
+            )
+        ]
+
+
+def restore_speech_timestamps(
+    segments: Iterable[Segment],
+    speech_chunks: List[dict],
+    sampling_rate: int,
+) -> Iterable[Segment]:
+    ts_map = SpeechTimestampsMap(speech_chunks, sampling_rate)
+
+    for segment in segments:
+        if segment.words:
+            words = []
+            for word in segment.words:
+                # Ensure the word start and end times are resolved to the same chunk.
+                middle = (word.start + word.end) / 2
+                chunk_index = ts_map.get_chunk_index(middle)
+                word = word._replace(
+                    start=ts_map.get_original_time(word.start, chunk_index),
+                    end=ts_map.get_original_time(word.end, chunk_index),
+                )
+                words.append(word)
+
+            segment = segment._replace(
+                start=words[0].start,
+                end=words[-1].end,
+                words=words,
+            )
+
+        else:
+            segment = segment._replace(
+                start=ts_map.get_original_time(segment.start),
+                end=ts_map.get_original_time(segment.end),
+            )
+
+        yield segment
+
+
+def get_ctranslate2_storage(segment: np.ndarray) -> ctranslate2.StorageView:
+    segment = np.ascontiguousarray(segment)
+    segment = ctranslate2.StorageView.from_array(segment)
+    return segment
+
+
+def get_compression_ratio(text: str) -> float:
+    text_bytes = text.encode("utf-8")
+    return len(text_bytes) / len(zlib.compress(text_bytes))
+
+
+def get_suppressed_tokens(
+    tokenizer: Tokenizer,
+    suppress_tokens: Optional[List[int]],
+) -> Optional[List[int]]:
+    if not suppress_tokens or -1 in suppress_tokens:
+        return suppress_tokens
+
+    suppress_tokens = list(suppress_tokens)
+
+    # Ensure the following special tokens are suppressed when the user does
+    # not use the default set (-1).
+    suppress_tokens.extend(
+        [
+            tokenizer.transcribe,
+            tokenizer.translate,
+            tokenizer.sot,
+            tokenizer.sot_prev,
+            tokenizer.sot_lm,
+        ]
+    )
+
+    return sorted(set(suppress_tokens))
+
+
+def merge_punctuations(alignment: List[dict], prepended: str, appended: str) -> None:
+    # merge prepended punctuations
+    i = len(alignment) - 2
+    j = len(alignment) - 1
+    while i >= 0:
+        previous = alignment[i]
+        following = alignment[j]
+        if previous["word"].startswith(" ") and previous["word"].strip() in prepended:
+            # prepend it to the following word
+            following["word"] = previous["word"] + following["word"]
+            following["tokens"] = previous["tokens"] + following["tokens"]
+            previous["word"] = ""
+            previous["tokens"] = []
+        else:
+            j = i
+        i -= 1
+
+    # merge appended punctuations
+    i = 0
+    j = 1
+    while j < len(alignment):
+        previous = alignment[i]
+        following = alignment[j]
+        if not previous["word"].endswith(" ") and following["word"] in appended:
+            # append it to the previous word
+            previous["word"] = previous["word"] + following["word"]
+            previous["tokens"] = previous["tokens"] + following["tokens"]
+            following["word"] = ""
+            following["tokens"] = []
+        else:
+            i = j
+        j += 1

utils.py

@@ -0,0 +1,157 @@
+import logging
+import os
+import re
+
+from typing import List, Optional
+
+import huggingface_hub
+import requests
+
+from tqdm.auto import tqdm
+
+_MODELS = {
+    "tiny.en": "Systran/faster-whisper-tiny.en",
+    "tiny": "Systran/faster-whisper-tiny",
+    "base.en": "Systran/faster-whisper-base.en",
+    "base": "Systran/faster-whisper-base",
+    "small.en": "Systran/faster-whisper-small.en",
+    "small": "Systran/faster-whisper-small",
+    "medium.en": "Systran/faster-whisper-medium.en",
+    "medium": "Systran/faster-whisper-medium",
+    "large-v1": "Systran/faster-whisper-large-v1",
+    "large-v2": "Systran/faster-whisper-large-v2",
+    "large-v3": "Systran/faster-whisper-large-v3",
+    "large": "Systran/faster-whisper-large-v3",
+    "distil-large-v2": "Systran/faster-distil-whisper-large-v2",
+    "distil-medium.en": "Systran/faster-distil-whisper-medium.en",
+    "distil-small.en": "Systran/faster-distil-whisper-small.en",
+    "distil-large-v3": "Systran/faster-distil-whisper-large-v3",
+}
+
+
+def available_models() -> List[str]:
+    """Returns the names of available models."""
+    return list(_MODELS.keys())
+
+
+def get_assets_path():
+    """Returns the path to the assets directory."""
+    return os.path.join(os.path.dirname(os.path.abspath(__file__)), "assets")
+
+
+def get_logger():
+    """Returns the module logger."""
+    return logging.getLogger("faster_whisper")
+
+
+def download_model(
+    size_or_id: str,
+    output_dir: Optional[str] = None,
+    local_files_only: bool = False,
+    cache_dir: Optional[str] = None,
+):
+    """Downloads a CTranslate2 Whisper model from the Hugging Face Hub.
+
+    Args:
+      size_or_id: Size of the model to download from https://huggingface.co/Systran
+        (tiny, tiny.en, base, base.en, small, small.en, distil-small.en, medium, medium.en,
+        distil-medium.en, large-v1, large-v2, large-v3, large, distil-large-v2,
+        distil-large-v3), or a CTranslate2-converted model ID from the Hugging Face Hub
+        (e.g. Systran/faster-whisper-large-v3).
+      output_dir: Directory where the model should be saved. If not set, the model is saved in
+        the cache directory.
+      local_files_only:  If True, avoid downloading the file and return the path to the local
+        cached file if it exists.
+      cache_dir: Path to the folder where cached files are stored.
+
+    Returns:
+      The path to the downloaded model.
+
+    Raises:
+      ValueError: if the model size is invalid.
+    """
+    if re.match(r".*/.*", size_or_id):
+        repo_id = size_or_id
+    else:
+        repo_id = _MODELS.get(size_or_id)
+        if repo_id is None:
+            raise ValueError(
+                "Invalid model size '%s', expected one of: %s"
+                % (size_or_id, ", ".join(_MODELS.keys()))
+            )
+
+    allow_patterns = [
+        "config.json",
+        "preprocessor_config.json",
+        "model.bin",
+        "tokenizer.json",
+        "vocabulary.*",
+    ]
+
+    kwargs = {
+        "local_files_only": local_files_only,
+        "allow_patterns": allow_patterns,
+        "tqdm_class": disabled_tqdm,
+    }
+
+    if output_dir is not None:
+        kwargs["local_dir"] = output_dir
+        kwargs["local_dir_use_symlinks"] = False
+
+    if cache_dir is not None:
+        kwargs["cache_dir"] = cache_dir
+
+    try:
+        return huggingface_hub.snapshot_download(repo_id, **kwargs)
+    except (
+        huggingface_hub.utils.HfHubHTTPError,
+        requests.exceptions.ConnectionError,
+    ) as exception:
+        logger = get_logger()
+        logger.warning(
+            "An error occured while synchronizing the model %s from the Hugging Face Hub:\n%s",
+            repo_id,
+            exception,
+        )
+        logger.warning(
+            "Trying to load the model directly from the local cache, if it exists."
+        )
+
+        kwargs["local_files_only"] = True
+        return huggingface_hub.snapshot_download(repo_id, **kwargs)
+
+
+def format_timestamp(
+    seconds: float,
+    always_include_hours: bool = False,
+    decimal_marker: str = ".",
+) -> str:
+    assert seconds >= 0, "non-negative timestamp expected"
+    milliseconds = round(seconds * 1000.0)
+
+    hours = milliseconds // 3_600_000
+    milliseconds -= hours * 3_600_000
+
+    minutes = milliseconds // 60_000
+    milliseconds -= minutes * 60_000
+
+    seconds = milliseconds // 1_000
+    milliseconds -= seconds * 1_000
+
+    hours_marker = f"{hours:02d}:" if always_include_hours or hours > 0 else ""
+    return (
+        f"{hours_marker}{minutes:02d}:{seconds:02d}{decimal_marker}{milliseconds:03d}"
+    )
+
+
+class disabled_tqdm(tqdm):
+    def __init__(self, *args, **kwargs):
+        kwargs["disable"] = True
+        super().__init__(*args, **kwargs)
+
+
+def get_end(segments: List[dict]) -> Optional[float]:
+    return next(
+        (w["end"] for s in reversed(segments) for w in reversed(s["words"])),
+        segments[-1]["end"] if segments else None,
+    )

vad.py

@@ -0,0 +1,291 @@
+import bisect
+import functools
+import os
+import warnings
+
+from typing import List, NamedTuple, Optional
+
+import numpy as np
+
+from faster_whisper.utils import get_assets_path
+
+
+# The code below is adapted from https://github.com/snakers4/silero-vad.
+class VadOptions(NamedTuple):
+    """VAD options.
+
+    Attributes:
+      threshold: Speech threshold. Silero VAD outputs speech probabilities for each audio chunk,
+        probabilities ABOVE this value are considered as SPEECH. It is better to tune this
+        parameter for each dataset separately, but "lazy" 0.5 is pretty good for most datasets.
+      min_speech_duration_ms: Final speech chunks shorter min_speech_duration_ms are thrown out.
+      max_speech_duration_s: Maximum duration of speech chunks in seconds. Chunks longer
+        than max_speech_duration_s will be split at the timestamp of the last silence that
+        lasts more than 100ms (if any), to prevent aggressive cutting. Otherwise, they will be
+        split aggressively just before max_speech_duration_s.
+      min_silence_duration_ms: In the end of each speech chunk wait for min_silence_duration_ms
+        before separating it
+      window_size_samples: Audio chunks of window_size_samples size are fed to the silero VAD model.
+        WARNING! Silero VAD models were trained using 512, 1024, 1536 samples for 16000 sample rate.
+        Values other than these may affect model performance!!
+      speech_pad_ms: Final speech chunks are padded by speech_pad_ms each side
+    """
+
+    threshold: float = 0.5
+    min_speech_duration_ms: int = 250
+    max_speech_duration_s: float = float("inf")
+    min_silence_duration_ms: int = 2000
+    window_size_samples: int = 1024
+    speech_pad_ms: int = 400
+
+
+def get_speech_timestamps(
+    audio: np.ndarray,
+    vad_options: Optional[VadOptions] = None,
+    **kwargs,
+) -> List[dict]:
+    """This method is used for splitting long audios into speech chunks using silero VAD.
+
+    Args:
+      audio: One dimensional float array.
+      vad_options: Options for VAD processing.
+      kwargs: VAD options passed as keyword arguments for backward compatibility.
+
+    Returns:
+      List of dicts containing begin and end samples of each speech chunk.
+    """
+    if vad_options is None:
+        vad_options = VadOptions(**kwargs)
+
+    threshold = vad_options.threshold
+    min_speech_duration_ms = vad_options.min_speech_duration_ms
+    max_speech_duration_s = vad_options.max_speech_duration_s
+    min_silence_duration_ms = vad_options.min_silence_duration_ms
+    window_size_samples = vad_options.window_size_samples
+    speech_pad_ms = vad_options.speech_pad_ms
+
+    if window_size_samples not in [512, 1024, 1536]:
+        warnings.warn(
+            "Unusual window_size_samples! Supported window_size_samples:\n"
+            " - [512, 1024, 1536] for 16000 sampling_rate"
+        )
+
+    sampling_rate = 16000
+    min_speech_samples = sampling_rate * min_speech_duration_ms / 1000
+    speech_pad_samples = sampling_rate * speech_pad_ms / 1000
+    max_speech_samples = (
+        sampling_rate * max_speech_duration_s
+        - window_size_samples
+        - 2 * speech_pad_samples
+    )
+    min_silence_samples = sampling_rate * min_silence_duration_ms / 1000
+    min_silence_samples_at_max_speech = sampling_rate * 98 / 1000
+
+    audio_length_samples = len(audio)
+
+    model = get_vad_model()
+    state = model.get_initial_state(batch_size=1)
+
+    speech_probs = []
+    for current_start_sample in range(0, audio_length_samples, window_size_samples):
+        chunk = audio[current_start_sample : current_start_sample + window_size_samples]
+        if len(chunk) < window_size_samples:
+            chunk = np.pad(chunk, (0, int(window_size_samples - len(chunk))))
+        speech_prob, state = model(chunk, state, sampling_rate)
+        speech_probs.append(speech_prob)
+
+    triggered = False
+    speeches = []
+    current_speech = {}
+    neg_threshold = threshold - 0.15
+
+    # to save potential segment end (and tolerate some silence)
+    temp_end = 0
+    # to save potential segment limits in case of maximum segment size reached
+    prev_end = next_start = 0
+
+    for i, speech_prob in enumerate(speech_probs):
+        if (speech_prob >= threshold) and temp_end:
+            temp_end = 0
+            if next_start < prev_end:
+                next_start = window_size_samples * i
+
+        if (speech_prob >= threshold) and not triggered:
+            triggered = True
+            current_speech["start"] = window_size_samples * i
+            continue
+
+        if (
+            triggered
+            and (window_size_samples * i) - current_speech["start"] > max_speech_samples
+        ):
+            if prev_end:
+                current_speech["end"] = prev_end
+                speeches.append(current_speech)
+                current_speech = {}
+                # previously reached silence (< neg_thres) and is still not speech (< thres)
+                if next_start < prev_end:
+                    triggered = False
+                else:
+                    current_speech["start"] = next_start
+                prev_end = next_start = temp_end = 0
+            else:
+                current_speech["end"] = window_size_samples * i
+                speeches.append(current_speech)
+                current_speech = {}
+                prev_end = next_start = temp_end = 0
+                triggered = False
+                continue
+
+        if (speech_prob < neg_threshold) and triggered:
+            if not temp_end:
+                temp_end = window_size_samples * i
+            # condition to avoid cutting in very short silence
+            if (window_size_samples * i) - temp_end > min_silence_samples_at_max_speech:
+                prev_end = temp_end
+            if (window_size_samples * i) - temp_end < min_silence_samples:
+                continue
+            else:
+                current_speech["end"] = temp_end
+                if (
+                    current_speech["end"] - current_speech["start"]
+                ) > min_speech_samples:
+                    speeches.append(current_speech)
+                current_speech = {}
+                prev_end = next_start = temp_end = 0
+                triggered = False
+                continue
+
+    if (
+        current_speech
+        and (audio_length_samples - current_speech["start"]) > min_speech_samples
+    ):
+        current_speech["end"] = audio_length_samples
+        speeches.append(current_speech)
+
+    for i, speech in enumerate(speeches):
+        if i == 0:
+            speech["start"] = int(max(0, speech["start"] - speech_pad_samples))
+        if i != len(speeches) - 1:
+            silence_duration = speeches[i + 1]["start"] - speech["end"]
+            if silence_duration < 2 * speech_pad_samples:
+                speech["end"] += int(silence_duration // 2)
+                speeches[i + 1]["start"] = int(
+                    max(0, speeches[i + 1]["start"] - silence_duration // 2)
+                )
+            else:
+                speech["end"] = int(
+                    min(audio_length_samples, speech["end"] + speech_pad_samples)
+                )
+                speeches[i + 1]["start"] = int(
+                    max(0, speeches[i + 1]["start"] - speech_pad_samples)
+                )
+        else:
+            speech["end"] = int(
+                min(audio_length_samples, speech["end"] + speech_pad_samples)
+            )
+
+    return speeches
+
+
+def collect_chunks(audio: np.ndarray, chunks: List[dict]) -> np.ndarray:
+    """Collects and concatenates audio chunks."""
+    if not chunks:
+        return np.array([], dtype=np.float32)
+
+    return np.concatenate([audio[chunk["start"] : chunk["end"]] for chunk in chunks])
+
+
+class SpeechTimestampsMap:
+    """Helper class to restore original speech timestamps."""
+
+    def __init__(self, chunks: List[dict], sampling_rate: int, time_precision: int = 2):
+        self.sampling_rate = sampling_rate
+        self.time_precision = time_precision
+        self.chunk_end_sample = []
+        self.total_silence_before = []
+
+        previous_end = 0
+        silent_samples = 0
+
+        for chunk in chunks:
+            silent_samples += chunk["start"] - previous_end
+            previous_end = chunk["end"]
+
+            self.chunk_end_sample.append(chunk["end"] - silent_samples)
+            self.total_silence_before.append(silent_samples / sampling_rate)
+
+    def get_original_time(
+        self,
+        time: float,
+        chunk_index: Optional[int] = None,
+    ) -> float:
+        if chunk_index is None:
+            chunk_index = self.get_chunk_index(time)
+
+        total_silence_before = self.total_silence_before[chunk_index]
+        return round(total_silence_before + time, self.time_precision)
+
+    def get_chunk_index(self, time: float) -> int:
+        sample = int(time * self.sampling_rate)
+        return min(
+            bisect.bisect(self.chunk_end_sample, sample),
+            len(self.chunk_end_sample) - 1,
+        )
+
+
+@functools.lru_cache
+def get_vad_model():
+    """Returns the VAD model instance."""
+    path = os.path.join(get_assets_path(), "silero_vad.onnx")
+    return SileroVADModel(path)
+
+
+class SileroVADModel:
+    def __init__(self, path):
+        try:
+            import onnxruntime
+        except ImportError as e:
+            raise RuntimeError(
+                "Applying the VAD filter requires the onnxruntime package"
+            ) from e
+
+        opts = onnxruntime.SessionOptions()
+        opts.inter_op_num_threads = 1
+        opts.intra_op_num_threads = 1
+        opts.log_severity_level = 4
+
+        self.session = onnxruntime.InferenceSession(
+            path,
+            providers=["CPUExecutionProvider"],
+            sess_options=opts,
+        )
+
+    def get_initial_state(self, batch_size: int):
+        h = np.zeros((2, batch_size, 64), dtype=np.float32)
+        c = np.zeros((2, batch_size, 64), dtype=np.float32)
+        return h, c
+
+    def __call__(self, x, state, sr: int):
+        if len(x.shape) == 1:
+            x = np.expand_dims(x, 0)
+        if len(x.shape) > 2:
+            raise ValueError(
+                f"Too many dimensions for input audio chunk {len(x.shape)}"
+            )
+        if sr / x.shape[1] > 31.25:
+            raise ValueError("Input audio chunk is too short")
+
+        h, c = state
+
+        ort_inputs = {
+            "input": x,
+            "h": h,
+            "c": c,
+            "sr": np.array(sr, dtype="int64"),
+        }
+
+        out, h, c = self.session.run(None, ort_inputs)
+        state = (h, c)
+
+        return out, state

version.py

@@ -0,0 +1,3 @@
+"""Version information."""
+
+__version__ = "1.0.2"