---

Hypa-LibreSpeech

Hypa-LibreSpeech is a large-scale, multilingual speech dataset curated by Hypa AI for training and evaluating Automatic Speech Recognition (ASR) and Text-to-Speech (TTS) systems across 8 European languages. It contains 200,000 high-quality audio–text pairs derived from open-domain audiobook recordings originally sourced from the LibriVox project.

This dataset builds directly upon two foundational open-source corpora:

• openslr/librispeech_asr — the original English LibriSpeech corpus by Panayotov et al. (2015)
• facebook/multilingual_librispeech — the Multilingual LibriSpeech (MLS) corpus by Pratap et al. (2020), covering 8 languages derived from LibriVox audiobooks

Hypa-LibreSpeech extends and repackages these sources into a unified, streamable, Whisper-ready format designed for modern multilingual speech model training.

This dataset is part of the broader Hypa AI open data initiative, which aims to democratize access to high-quality speech data for AI researchers and developers worldwide.

---

Dataset Summary

Hypa-LibreSpeech is an open-source multilingual speech dataset derived from the LibreSpeech corpus and curated to support the development of robust speech and language technologies across multiple languages. The dataset contains approximately 200,000 high-quality audio–text pairs spanning 8 languages, providing a diverse collection of speech recordings from multiple speakers, accents, and speaking styles.

The dataset includes aligned speech recordings and text transcriptions, making it suitable for a wide range of speech processing and language modeling tasks. Audio samples range from 1.3 to 22.5 seconds in duration, while transcription lengths vary from 4 to 565 characters, capturing both short utterances and longer spoken passages. To support different deployment scenarios, audio is provided in both FLAC (lossless) and OPUS (compressed) formats.

This release provides:

Multilingual Speech–Text Pairs: Paired audio recordings and transcriptions across 8 languages. Multiple Audio Formats: Speech data available in FLAC and OPUS formats for training and deployment flexibility. Diverse Speaker Coverage: Recordings from a broad pool of speakers with varying accents and speaking characteristics. Unified Transcription Task: Standardized speech-to-text samples designed for multilingual automatic speech recognition research.

The dataset is designed to support a variety of downstream tasks, including:

Automatic Speech Recognition (ASR): Fine-tuning and evaluating multilingual ASR models such as Whisper, Wav2Vec2, and MMS. Text-to-Speech (TTS): Training multilingual and cross-lingual speech synthesis systems such as XTTS, VITS, and Orpheus. Speech Translation Research: Building and evaluating multilingual speech processing pipelines. Cross-Lingual Transfer Learning: Investigating knowledge transfer between high-resource and low-resource languages. Speech Representation Learning: Pretraining and benchmarking speech foundation models. Language Identification and Classification: Developing systems for spoken language recognition and multilingual speech understanding.

By providing a large-scale multilingual collection of aligned speech and text data, Hypa-LibreSpeech aims to facilitate research and development of inclusive speech technologies that generalize effectively across languages and speaker populations.

---

Supported Tasks and Leaderboards

Task	Description
automatic-speech-recognition	Transcribe spoken audio to text in the source language
text-to-speech	Use paired text–audio for voice synthesis training
audio-classification	Language identification from audio

---

Languages

The dataset covers 8 European languages, matching the language scope of the upstream MLS corpus:

Language	ISO Code	Script
English	en	Latin
French	fr	Latin
German	de	Latin
Dutch	nl	Latin
Spanish	es	Latin
Italian	it	Latin
Portuguese	pt	Latin
Polish	pl	Latin

---

Dataset Structure

Data Instances

Each instance in the dataset represents a single utterance and contains the following:

{
  "audio": {
    "path": "audio/train/0001.opus",
    "array": [...],          # decoded audio waveform
    "sampling_rate": 16000
  },
  "text": "ook in de allerslechtste dat meestal mijn geval is...",
  "src_lang": "dutch",
  "tgt_lang": "dutch",
  "duration_seconds": 11.12,
  "mode": "transcribe",
  "speaker": "random"
}

Data Fields

Field	Type	Description
audio	Audio	Audio object containing the waveform array, path, and sampling rate (16,000 Hz)
text	string	Ground-truth text transcription of the audio clip
src_lang	string	Source language of the audio (e.g., "dutch", "english")
tgt_lang	string	Target language of the transcription (same as src_lang for transcription tasks)
duration_seconds	float	Duration of the audio clip in seconds
mode	string	Task mode — currently "transcribe" for all instances
speaker	string	Speaker label — currently "random", reflecting the diverse volunteer speaker pool

Data Splits

Split	Num. Examples
train	200,000

The dataset ships as a single train split. Users are encouraged to create their own validation and test splits as needed. For reference, the upstream MLS corpus provides standardized dev and test splits per language that can be used for evaluation.

---

Dataset Creation

Source Data

Hypa-LibreSpeech is derived from two primary upstream corpora:

1. LibriSpeech ASR Corpus

Panayotov, V., Chen, G., Povey, D., & Khudanpur, S. (2015). LibriSpeech: An ASR corpus based on public domain audio books. ICASSP.

Available on Hugging Face as openslr/librispeech_asr.

LibriSpeech is a corpus of approximately 1,000 hours of 16kHz read English speech, prepared by Vassil Panayotov with the assistance of Daniel Povey. It is derived from read audiobooks from the LibriVox project, and has been carefully segmented and aligned.

2. Multilingual LibriSpeech (MLS)

Pratap, V., Xu, Q., Sriram, A., Synnaeve, G., & Collobert, R. (2020). MLS: A Large-Scale Multilingual Dataset for Speech Research. ArXiv:2012.03411.

Available on Hugging Face as facebook/multilingual_librispeech.

MLS is a large multilingual corpus derived from read audiobooks from LibriVox, covering 8 languages: English, German, Dutch, Spanish, French, Italian, Portuguese, and Polish. It includes approximately 44,500 hours of English and a total of approximately 6,000 hours for the other 7 languages, making it one of the largest publicly available multilingual speech datasets.

3. LibriVox (Original Audio Source)

All audio ultimately originates from LibriVox, a volunteer-driven project that produces free public domain audiobook recordings. LibriVox releases all recordings under the CC0 1.0 Universal (Public Domain) license.

Curation Rationale

While MLS and LibriSpeech are foundational resources, they are not always straightforward to use directly for modern training pipelines — particularly for fine-tuning instruction-following speech models like Whisper. Hypa-LibreSpeech addresses this by:

• Unifying schema: Providing a single consistent schema across all 8 languages with src_lang, tgt_lang, mode, and duration_seconds fields
• Filtering for quality: Removing segments with transcription alignment issues, excessively short clips (< 1.3s), or very long clips (> 22.5s)
• Optimizing for streaming: Encoding audio in OPUS format for efficient streaming alongside lossless FLAC archives
• Repackaging for modern use: Structuring data to be immediately usable with HuggingFace datasets, transformers, and common training frameworks

Processing Pipeline

The dataset was produced through the following pipeline:

1. Source ingestion — Audio and aligned text were ingested from the MLS and LibriSpeech corpora via the Hugging Face Datasets library.
2. Segmentation validation — Segment boundaries were validated against the original LibriVox source timestamps. Segments with misalignments were discarded.
3. Duration filtering — Clips shorter than 1.3s or longer than 22.5s were removed to ensure training stability.
4. Text normalization — Transcriptions were lowercased and lightly normalized (removing chapter headers, annotations, and non-speech markers).
5. Audio re-encoding — Segments were re-encoded to 16 kHz mono in both FLAC (lossless) and OPUS (compressed, ~50% size reduction).
6. Language tagging — Each segment was tagged with src_lang and tgt_lang derived from the originating corpus language metadata.
7. Schema standardization — All fields were aligned to the unified Hypa-LibreSpeech schema and exported to Parquet for efficient loading.

---

Statistics

Metric	Value
Total Examples	200,000
Total Languages	8
Min Duration	1.3 seconds
Max Duration	22.5 seconds
Estimated Total Audio	~600 hours
Audio Formats	FLAC, OPUS
Sampling Rate	16,000 Hz
Text Min Length	4 characters
Text Max Length	565 characters
Speaker Pool	Mixed (volunteer readers via LibriVox)
Task Mode	transcribe

Approximate Language Distribution

Language	Approx. Examples	Source
English	~25,000	LibriSpeech + MLS
French	~25,000	MLS
German	~25,000	MLS
Dutch	~25,000	MLS
Spanish	~25,000	MLS
Italian	~25,000	MLS
Portuguese	~25,000	MLS
Polish	~25,000	MLS

For reference, the full upstream MLS corpus contains the following training hours per language:

Language	MLS Train Hours	MLS Train Samples
English	44,659	—
German	1,967	469,942
Dutch	1,554	374,287
French	1,077	258,213
Spanish	918	220,701
Italian	247	59,623
Portuguese	161	37,533
Polish	104	25,043

Hypa-LibreSpeech draws a balanced 200k-sample subset from these sources.

---

Usage

Loading the Dataset

from datasets import load_dataset

# Load the full dataset
dataset = load_dataset("hypaai/Hypa-LibreSpeech")

# View a sample
print(dataset["train"][0])

Filtering by Language

from datasets import load_dataset

dataset = load_dataset("hypaai/Hypa-LibreSpeech", split="train")

# Filter for French examples only
french_data = dataset.filter(lambda x: x["src_lang"] == "french")
print(f"French examples: {len(french_data)}")

Fine-tuning Whisper

from datasets import load_dataset
from transformers import WhisperForConditionalGeneration, WhisperProcessor

# Load dataset
dataset = load_dataset("hypaai/Hypa-LibreSpeech", split="train")

# Load Whisper processor and model
processor = WhisperProcessor.from_pretrained("openai/whisper-small")
model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-small")

# Prepare a single sample
sample = dataset[0]
inputs = processor(
    sample["audio"]["array"],
    sampling_rate=sample["audio"]["sampling_rate"],
    return_tensors="pt"
)

# Generate transcription
predicted_ids = model.generate(inputs["input_features"])
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)
print(transcription)

Streaming for Large-Scale Training

from datasets import load_dataset

# Stream without downloading the full dataset
dataset = load_dataset("hypaai/Hypa-LibreSpeech", split="train", streaming=True)

for sample in dataset.take(5):
    print(sample["text"], "|", sample["src_lang"], "|", sample["duration_seconds"], "s")

Using with PyTorch DataLoader

from datasets import load_dataset
from torch.utils.data import DataLoader

dataset = load_dataset("hypaai/Hypa-LibreSpeech", split="train", streaming=True)
dataloader = DataLoader(dataset, batch_size=16)

---

Considerations for Using the Data

Social Impact

Hypa-LibreSpeech contributes to the democratization of multilingual speech AI by making a large, well-structured, open-access speech corpus freely available. By lowering the barrier to building ASR and TTS systems across 8 European languages, it supports researchers, startups, and developers who may not have the resources to curate their own large-scale training data.

Biases

• Speaker demographics: Audio originates from LibriVox volunteer readers, who skew toward native or near-native speakers and do not uniformly represent all regional accents, dialects, or age groups. Refer to the MLS paper (Pratap et al., 2020) for detailed speaker gender statistics per language.
• Domain bias: All text is from literary/audiobook sources (19th–early 20th century literature). This differs significantly from conversational, spontaneous, or technical domain speech. Models trained exclusively on this data may underperform on informal or domain-specific audio.
• Language imbalance: The upstream MLS corpus has significant imbalance in total training hours across languages (English: ~44k hours vs. Polish: ~104 hours). Hypa-LibreSpeech rebalances this with a ~25k sample cap per language, but downstream model performance may still vary.
• Text style: Transcriptions reflect the literary register of the source texts, including complex sentence structures and archaic vocabulary in some languages.

Limitations

• The dataset contains a single train split. Users must create their own validation and test sets, or use the standardized splits from the upstream facebook/multilingual_librispeech dataset for evaluation.
• The speaker field is set to "random" for all entries. Speaker-level diarization or attribution is not included in this release.
• Audio quality varies naturally due to the volunteer recording nature of LibriVox — some recordings contain minor background noise, room echo, or microphone variation.
• The dataset does not include timestamps at the word or phoneme level.

---

Licensing Information

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license, consistent with the license of the upstream MLS corpus.

The underlying audio recordings are sourced from LibriVox, released into the public domain under the CC0 1.0 Universal license.

You are free to use, share, and adapt this dataset for any purpose, including commercial use, provided appropriate attribution is given to Hypa AI and the upstream sources.

---

Citation Information

If you use Hypa-LibreSpeech in your research or projects, please cite this dataset and the upstream corpora it is derived from:

Cite Hypa-LibreSpeech

@dataset{hypaai2025librespeech,
  title        = {Hypa-LibreSpeech: A Multilingual Audiobook Speech Dataset},
  author       = {Hypa AI},
  year         = {2025},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/hypaai/Hypa-LibreSpeech},
  license      = {CC BY 4.0},
  note         = {200,000 audio-text pairs across 8 European languages, derived from LibriSpeech and Multilingual LibriSpeech (MLS).}
}

Cite Multilingual LibriSpeech (MLS)

@article{Pratap2020MLSAL,
  title   = {MLS: A Large-Scale Multilingual Dataset for Speech Research},
  author  = {Vineel Pratap and Qiantong Xu and Anuroop Sriram and Gabriel Synnaeve and Ronan Collobert},
  journal = {ArXiv},
  year    = {2020},
  volume  = {abs/2012.03411},
  url     = {https://arxiv.org/abs/2012.03411}
}

Cite LibriSpeech

@inproceedings{panayotov2015librispeech,
  title     = {LibriSpeech: An ASR corpus based on public domain audio books},
  author    = {Panayotov, Vassil and Chen, Guoguo and Povey, Daniel and Khudanpur, Sanjeev},
  booktitle = {2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages     = {5206--5210},
  year      = {2015},
  organization = {IEEE}
}

Acknowledge LibriVox

LibriVox (https://librivox.org) — free public domain audiobooks read by volunteers worldwide.

---

Contributions

This dataset was created and maintained by the Hypa AI team.

We welcome contributions, bug reports, and discussion through the Community tab.

If you build models or downstream datasets using Hypa-LibreSpeech, we'd love to hear about it — please tag us or open a discussion.

---

Part of the Hypa AI open data initiative. See also: Hypa_Fleurs