app.py

import os, torch, gradio as gr
from transformers import AutoProcessor, AutoModelForCTC

# --- CONFIG ---
MODEL_ID = os.getenv("MODEL_ID", "Reihaneh/wav2vec2_fy_nl_best_frisian_1")
HF_TOKEN = os.getenv("HF_TOKEN")  # only if the repo is private



# Faster hub downloads (optional but nice if you also add the package 'hf-transfer')
os.environ.setdefault("HF_HUB_ENABLE_HF_TRANSFER", "1")

device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.float16 if device == "cuda" else torch.float32

# --- LOAD ONCE (local cache persists across Space restarts) ---
processor = AutoProcessor.from_pretrained(MODEL_ID, token=HF_TOKEN)
model = AutoModelForCTC.from_pretrained(MODEL_ID, token=HF_TOKEN, torch_dtype=dtype)
model.to(device).eval()

target_sr = getattr(getattr(processor, "feature_extractor", None), "sampling_rate", 16000)

def transcribe(audio):
    if audio is None:
        return ""
    sr, wav = audio  # gr.Audio(type="numpy") returns (sr, np.ndarray)
    if wav.ndim == 2:  # mono
        wav = wav[:, 0]
    # quick resample if needed (basic, avoids extra deps)
    if sr != target_sr:
        # naive linear resample to avoid importing librosa
        import numpy as np
        import math
        ratio = target_sr / sr
        idx = (np.arange(int(math.ceil(wav.shape[0] * ratio))) / ratio).astype(int)
        idx = idx[idx < wav.shape[0]]
        wav = wav[idx]

    inputs = processor(wav, sampling_rate=target_sr, return_tensors="pt")
    with torch.inference_mode():
        logits = model(inputs.input_values.to(device)).logits
        ids = torch.argmax(logits, dim=-1)
        text = processor.batch_decode(ids)[0]
    return text

with gr.Blocks(title="Frisian ASR") as demo:
    gr.Markdown("## 🎙️ Frisian ASR\nUpload audio or use your mic.")
    audio = gr.Audio(sources=["microphone", "upload"], type="numpy", label="Audio")
    out = gr.Textbox(label="Transcript")
    gr.Button("Transcribe").click(transcribe, inputs=audio, outputs=out)
    # optional: auto-run when audio changes
    # audio.change(transcribe, inputs=audio, outputs=out, every=0)

demo.queue(concurrency_count=1, max_size=16)
if __name__ == "__main__":
    demo.launch()



#import gradio as gr
#gr.Interface.load("models/Reihaneh/wav2vec2_fy_nl_best_frisian_1").launch()



#import gradio as gr

#gr.load("Reihaneh/wav2vec2_fy_nl_best_frisian_1").launch(share=True)
'''
import os, torch, gradio as gr
from transformers import pipeline
import torch

# optional: keep memory low on CPU Spaces
if not torch.cuda.is_available():
    torch.set_num_threads(max(1, os.cpu_count() // 2))

MODEL_ID = os.getenv("MODEL_ID", "Reihaneh/wav2vec2_fy_nl_best_frisian_1")  # set in Space Secrets or hardcode
HF_TOKEN = os.getenv("HF_TOKEN")  # only needed if model is private

asr = pipeline(
    task="automatic-speech-recognition",
    model=MODEL_ID,
    # common safe defaults; adjust for your model
    chunk_length_s=20,              # streaming-style chunking for long audio
    stride_length_s=(4, 2),
    device_map="auto",              # uses GPU if available
    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
    token=HF_TOKEN,                 # works for private models
)

def transcribe(audio, language=None):
    """
    audio comes as (sr, np.ndarray) from gr.Audio when type='numpy'.
    If your model expects specific sampling rate, set 'sampling_rate' below.
    """
    if audio is None:
        return ""
    result = asr(
        audio,
        generate_kwargs={"task": "transcribe"} if "whisper" in MODEL_ID.lower() else None,
        return_timestamps=False,
    )
    # result is either a dict with 'text' or a string depending on transformers version
    return result["text"] if isinstance(result, dict) and "text" in result else str(result)

with gr.Blocks(title="ASR Demo") as demo:
    gr.Markdown("# 🎙️ ASR Demo\nUpload a file or speak into your mic.")
    with gr.Row():
        audio = gr.Audio(sources=["microphone", "upload"], type="numpy", label="Audio")
        lang = gr.Dropdown(choices=[None,"fy-NL","nl-NL","en-US"], value=None, label="(Optional) Language hint")
    out = gr.Textbox(label="Transcription")
    btn = gr.Button("Transcribe")
    btn.click(fn=transcribe, inputs=[audio, lang], outputs=out)
    # optional: live/auto mode (non-streaming)
    audio.change(fn=transcribe, inputs=[audio, lang], outputs=out, every=0)

demo.queue(max_size=12, concurrency_count=1).launch()'''