Skip to content

About

📜 A list of various Audio/Speech datasets about Speech Recognition, Speech Synthesis, Noise, Audio Tagging/Sound Event Detection, Speaker Diarization, Speaker Recognition, (Inverse) Text normalization, Speech Translation, Multilingual, etc. (continuously update)

Topics

Resources

Stars

9 stars

Watchers

1 watching

Forks

Latest commit

 

History

18 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 

Repository files navigation

Audio/Speech Datasets

A list of various Audio/Speech datasets about Speech Recognition, Speech Synthesis, Noise, Audio Tagging/Sound Event Detection, Speaker Diarization, Speaker Recognition, Speaker and Speech Traits, (Inverse) Text normalization, Speech Translation, Multilingual, etc. (continuously update)

Table of contents generated with markdown-toc

Note

License / Terms: Official licenses are listed before additional usage terms. Not stated does not imply unrestricted use.

Access: Direct = direct download; Registration = account or contact information required; Gated = terms must be accepted; Application = access must be requested; Unavailable = no working official download was found. Labels may be combined when more than one requirement applies.

Metadata reviewed: 2026-09-02.

Task

Speech Recognition

Chinese

Dataset Hours License / Terms Access
THCHS-30 30 Apache-2.0 Direct
AISHELL-1 178 Apache-2.0 Direct
AISHELL-2 1,000 Academic use only Application
Free ST Chinese Mandarin (ST-CMDS) 110 CC BY-NC-ND 4.0 Direct
Primewords Chinese Corpus Set 1 99 CC BY-NC-ND 4.0 Direct
aidatatang_200zh 200 CC BY-NC-ND 4.0 Unavailable
aidatatang_1505zh 1,505 Academic use only Application
MAGICDATA Mandarin Read 755 CC BY-NC-ND 4.0 Direct
MAGICDATA Mandarin Conversational (RAMC) 180 CC BY-NC-ND 4.0 Direct
AliMeeting (M2MeT) 118.75 CC BY-SA 4.0 Direct
WenetSpeech 22,400+ CC BY 4.0 · non-commercial terms Application
TAL-ASR (Adult Chinese Teaching Audio) 100 Custom terms · internal research only Registration · Gated
DiDiSpeech 800 Not stated Unavailable

Note

  • aidatatang_200zh: The official download is currently unavailable; the entry and linked metadata/license record are retained.
  • AliMeeting: Also listed under Speaker Diarization. Its train/dev/test split is 104.75/4/10 hours.
  • WenetSpeech: Contains more than 22,400 hours in total, including more than 10,000 hours of high-quality labeled data. Its official page states CC BY 4.0 and separately limits the release to non-commercial use.
  • DiDiSpeech: Its former official download is no longer available; third-party mirrors are not treated as authoritative license sources.

English

Dataset Hours License / Terms Access
LibriSpeech 1,000 CC BY 4.0 Direct
GigaSpeech 33,000+ Apache-2.0 · non-commercial research/education terms Application
Libri-Light 60,000+ Not stated Direct
LibriHeavy 50,000 Apache-2.0 repository · audio terms not stated Direct
SPGISpeech 5,000 Custom terms · research use only Gated
The People's Speech 30,000+ CC BY 4.0 / CC BY-SA Direct

Note

  • GigaSpeech: Contains more than 33,000 hours in total, of which about 10,000 hours are transcribed; its audio remains subject to source-owner rights and the dataset's additional terms.
  • Libri-Light: Its repository license applies to the accompanying code; no explicit dataset license was found on the official dataset page.
  • LibriHeavy: Its Apache-2.0 license covers its repository; its audio is inherited from Libri-Light, whose official page does not state a dataset license.
  • The People's Speech: Its license varies by source subset; check the per-file metadata before reuse.

Multilingual

Dataset Hours License / Terms Access
Multilingual LibriSpeech (MLS) 50,000+ CC BY 4.0 Direct
TALCS / TAL-CSASR 587 Custom terms · internal research only Registration · Gated
FLEURS ~1,200 CC BY 4.0 Direct
YODAS 369,510 CC BY 3.0 Direct
VoxPopuli 1,791 transcribed CC0-1.0 · source legal notice Direct
Granary ~643,000 ASR CC BY 4.0 · CC BY 3.0 subset · source audio terms Direct
MOSEL 444,035 pseudo-labeled CC BY 4.0 annotations · source audio terms Direct

Note

  • TALCS / TAL-CSASR: Mandarin-English code-switching speech from online English lessons, with paired transcriptions. Also listed under Child/Adult Speech Resources as adult speech; per-utterance age and gender annotations have not been confirmed.
  • FLEURS: A 102-language read-speech benchmark built from 2,009 parallel FLoRes sentences, with approximately 12 hours per language and speaker-disjoint train/dev/test splits. It supports ASR, spoken language identification, and speech-to-text retrieval, and includes Mandarin Chinese and Cantonese. It is also available through TensorFlow Datasets.
  • YODAS: The linked release contains 369,510 hours of segmented 16 kHz YouTube speech with user-provided or automatically generated captions across 149 languages. Its manual label means that captions were uploaded by users, not necessarily transcribed by humans; treat the transcripts as weak labels and validate them before supervised training. YODAS2 packages the same underlying data as unsegmented, video-level 24 kHz audio with timestamped utterances, so its hours should not be added to YODAS. The wider YODAS project reports more than 500,000 hours when additional subsets are included.
  • VoxPopuli: The Hugging Face ASR release contains 1,791 transcribed hours across 16 core European languages, plus a 29-hour non-native English test set. The wider project also provides approximately 384,000 hours of unlabelled speech across 23 languages and 17,300 hours of interpretation data. Recordings come from European Parliament events; the dataset is CC0, while the raw source remains subject to the European Parliament legal notice.
  • Granary: Covers 25 European languages, with filtered, machine-generated transcriptions from YODAS2, VoxPopuli, YouTube-Commons, and Libri-Light. The main repository provides manifests; audio must be obtained separately using the download guide. The YODAS-Granary subset includes audio and uses CC BY 3.0; other source audio retains its own terms. Hours follow the official task summary, whose figures differ from the overview and corpus breakdown. Metadata reviewed: 2026-10-01.
  • MOSEL: The linked release lists 444,035 hours of automatic transcripts; the roughly 950,000-hour resource collection also includes existing labeled and unlabeled corpora. Initial transcripts used Whisper large-v3 on VoxPopuli and Libri-Light; v2 updates transcripts and adds YouTube-Commons data and English translations of non-English VoxPopuli speech. The release provides annotations, with audio obtained from the original sources. CC BY 4.0 applies to the published annotations; original audio terms still apply. Metadata reviewed: 2026-10-01.
  • Source overlap: Granary incorporates MOSEL annotations and reuses audio from corpora already listed here. Do not add Granary, MOSEL, YODAS, VoxPopuli, and Libri-Light hours together as independent audio; check source IDs and segments before combining training sets.

Speech Synthesis

Chinese

Dataset Hours License / Terms Access
AISHELL-3 85 Apache-2.0 Direct

English

Dataset Hours License / Terms Access
LibriTTS 585 CC BY 4.0 Direct

Noise

Dataset Hours License / Terms Access
MUSAN — CC BY 4.0 Direct
Aachen Impulse Response Database (AIR) — Not stated Direct
Simulated Room Impulse Response Database — Apache-2.0 Direct
Room Impulse Response and Noise Database — Apache-2.0 Direct

Audio Tagging/Sound Event Detection

Speaker Diarization

Dataset Hours License / Terms Access
AliMeeting (M2MeT) 118.75 CC BY-SA 4.0 Direct

Speaker Recognition

Speaker and Speech Traits

Attribute-annotated datasets

Dataset Hours License / Terms Access
TAL Adult Chinese Speech Emotion 12.5 Custom terms · internal research only Registration · Gated

Note

  • TAL Adult Chinese Speech Emotion: Contains 4,541 clips from 42 speakers. Each clip has Pleasure and Arousal scores in [-3, 3]; the official annotation example includes speaker ID and gender. This is a dimensional speech-emotion dataset.
  • TAL usage terms: Prohibit commercial use of the datasets and models trained from them, as well as redistribution or creation of derivative datasets.

Child/Adult Speech Resources

Dataset Hours License / Terms Access
TAL-ASR (Adult Chinese Teaching Audio) 100 Custom terms · internal research only Registration · Gated
TALCS / TAL-CSASR 587 Custom terms · internal research only Registration · Gated
TAL Child Chinese Read Speech 5.4 Custom terms · internal research only Registration · Gated
TAL Child English Read Speech 4.5 Custom terms · internal research only Registration · Gated
TAL Adult Chinese Read Speech 1,750 Custom terms · internal research only Registration · Gated
TAL Adult English Read Speech 180 Custom terms · internal research only Registration · Gated
TAL Adult English Teaching Audio 160 Custom terms · internal research only Registration · Gated

Note

  • Age/gender annotations: Child/adult groups come from the official corpus names. Exact ages and per-utterance age or gender annotations have not been confirmed; these are candidate audio resources for further annotation and validation.
  • TAL-ASR and TALCS: Also listed under Speech Recognition, with paired transcriptions. TAL-ASR explicitly includes speaker IDs, which do not establish age or gender labels.
  • Read speech and adult English teaching audio: The official page lists WAV audio without stating paired transcripts or per-utterance speaker/attribute labels. Verify the downloaded contents before treating these as supervised ASR, TTS, or attribute-recognition datasets.
  • TAL usage terms: Prohibit commercial use of the datasets and models trained from them, as well as redistribution or creation of derivative datasets.

(Inverse) Text normalization

Speech Translation

Dataset Hours License / Terms Access
GigaST 10,000 CC BY-NC 4.0 Direct
GigaS2S — CC BY 4.0 Direct
FLEURS ~1,200 CC BY 4.0 Direct
VoxPopuli 17,300 CC0-1.0 · source legal notice Direct
Granary ~351,000 AST CC BY 4.0 · CC BY 3.0 subset · source audio terms Direct

Note

  • GigaST: Provides English-to-German and English-to-Chinese translations; the source audio follows GigaSpeech's separate access terms.
  • GigaS2S: Provides English-to-Chinese synthetic target speech.
  • FLEURS: Records 2,009 parallel sentences in 102 languages, enabling multilingual speech-to-text translation evaluation. Its limited per-language scale makes it primarily a benchmark rather than a large translation training corpus.
  • VoxPopuli: Provides speech-to-speech interpretation data covering 15×15 language directions from European Parliament events; unlike GigaS2S, its target speech consists of real parliamentary interpretation rather than synthesized speech.
  • Granary: Provides machine-generated English text translations for speech in 24 non-English European languages, using EuroLLM and quality filtering. The target is text. ASR and translation entries reuse source audio, so their hours are not independent. Audio access and license details are described under Multilingual ASR. Metadata reviewed: 2026-10-01.

Reference

About

📜 A list of various Audio/Speech datasets about Speech Recognition, Speech Synthesis, Noise, Audio Tagging/Sound Event Detection, Speaker Diarization, Speaker Recognition, (Inverse) Text normalization, Speech Translation, Multilingual, etc. (continuously update)

Topics

Resources

Stars

9 stars

Watchers

1 watching

Forks

Releases

Packages

Contributors