15 Public Speech Datasets for ASR and TTS (2026)

For ASR, choose speech that matches the language and recording conditions you need to recognize. For TTS, check recordings, transcripts and speaker coverage against your model's requirements. These 15 public speech datasets cover different tasks, from short commands to meetings and read-speech synthesis.

This comparison groups corpora by intended task, recording style, published coverage and release terms. Figures were checked against the linked release information on 18 October 2026. They describe those releases, including older research corpora, rather than an amount of usable training audio guaranteed for a new project.

18 October 2026 revision: Updated Common Voice release figures and dataset license descriptions, corrected FLEURS training guidance, and removed unsupported rankings.

Compare the released data, language and task

The license column summarizes the linked distribution's stated terms. Read the actual release conditions and supporting documentation for your planned use, including separate conditions for code or models.

ASR data: distinguish reading, commands and conversation

1. LibriSpeech: read English with defined subsets

LibriSpeech derives from LibriVox audiobooks. Its three training subsets contain 100, 360 and 500 hours, with separate development and test downloads. That gives a clear starting structure for a read-English recognition experiment.

Keep the subset name beside a reported result. Before choosing it for calls or meetings, compare the reading style and recording conditions with your application. A result on one LibriSpeech test set answers a narrower question than performance on your own audio.

2. Common Voice: choose a modality, release and language

Mozilla's Scripted Speech v27 announcement reports 295 languages and about 42,593 total recorded hours. Spontaneous Speech v5 covers 80 languages and includes 302 hours that have been transcribed and validated. The first collection contains read prompts; the second contains free-form answers to questions.

These are different coverage measures. Inspect your language's release instead of treating a global total as available validated training audio. The Mozilla Data Collective listings identify versions, locales and CC0-1.0 terms for these releases. Follow the listing's download conditions as well as its license.

3. Multilingual LibriSpeech: audiobook speech in eight languages

MLS provides English, German, Dutch, Spanish, French, Italian, Portuguese and Polish read speech. Its original dataset paper describes about 44.5K English hours and about 6K hours across the other languages.

The release includes training, development and test sets, plus smaller training selections for limited-supervision experiments. Choose the language and subset explicitly. A combined multilingual total does not describe the amount available for each language or supply conversational recording conditions.

4. People's Speech: check the subset and its terms

MLCommons describes more than 30,000 hours of transcribed English speech, distributed in CC BY and CC BY-SA subsets. Its release provides a way to explore a larger English corpus with multiple source environments.

Record which subset you select and inspect transcript conventions on representative files. Review the license attached to that subset, the accompanying source information and the attribution requirements. Keep those details with the experiment so a later team member can identify the same data.

5. Speech Commands: isolated-word recognition

The original Speech Commands v2 paper reports 105,829 utterances of 35 words. Utterance files are one second or shorter, and the task is keyword spotting rather than full-sentence transcription. The paper also describes held-out files and evaluation procedures.

Use it when your question concerns recognizing short command words. For a continuously listening application, inspect how the chosen recipe handles unrelated words, silence and false triggers, as well as classification of the target words. Preserve the documented split files when reproducing an experiment.

6. VoxPopuli: separate unlabelled and transcribed speech

VoxPopuli's headline 400K hours are unlabelled speech across 23 languages. The official repository separately reports about 1.8K transcribed hours across 16 languages. The source recordings are European Parliament events from 2009 to 2020.

Choose the data type your training task requires before comparing volumes. The repository lists CC0 for VoxPopuli data and points to the Parliament's raw-data legal notice. It lists different terms for code and pretrained models, so keep those components separate in your release record.

7. TED-LIUM 3: use the correct distribution

TED-LIUM 3 contains English TED-talk speech. Its creator paper describes 452 hours and two distributions: a legacy arrangement and another designed for speaker-adaptation experiments. Preserve the matching audio, transcript and split versions when following a recipe.

TensorFlow's release-3 dataset entry documents CC BY-NC-ND 3.0. Include that restriction in your selection review rather than treating a public corpus listing as permission for every use. Check the current distribution instructions before planning a download.

8. GigaSpeech: choose the transcribed training subset

GigaSpeech v1.0.0 documents 10,000 hours of transcribed English drawn from audiobooks, podcasts and YouTube. Its larger total-audio figure is a different quantity. The repository provides several nested training subsets, together with separate development and test material.

Select a subset according to your experiment and inspect its source mix. Follow the dataset's documented access route. The repository's software license should be recorded separately from the conditions governing the recordings; a code license alone is not a dataset-use decision.

9. FLEURS: training and evaluation both have a role

Google's FLEURS card covers 102 languages and describes training speakers who differ from the development and test speakers. It explicitly supports multilingual fine-tuning. Calling it too small for training would exclude a documented use of the corpus.

Use the correct split for the question you are testing. Its read-speech focus also matters when interpreting results for noisier applications. Compare your application's speech and recording conditions separately rather than discarding the benchmark or assuming its score predicts deployment performance.

10. AMI: inspect the meeting and microphone choice

AMI contains 100 hours of meeting recordings, with close-talking and far-field microphones and orthographic transcription. Around two-thirds of the meetings were elicited through design-team scenarios; the remainder are naturally occurring meetings.

Specify the microphone channels, meeting subset and annotations you intend to use. The publisher's release statement covers signals, transcription and some annotations. Check the availability of a particular annotation layer rather than assuming every meeting has every label you need.

11. CHiME-6: identify the synchronized release

CHiME-6 provides multichannel far-field English speech and is derived from CHiME-5 with corrections to array synchronization. The current OpenSLR release lists CC BY-SA 4.0 and provides separate training, development, evaluation and transcription downloads.

For an experiment using distant microphones, record the release, channel selection and transcript version. Follow the relevant challenge protocol when reproducing challenge results. Historical challenge access instructions and the current dataset release should be checked in their own context.

12. AISHELL-1: read Mandarin with speaker information

OpenSLR's AISHELL-1 entry describes 400 speakers recording Mandarin in a quiet indoor environment and lists Apache 2.0. The original paper describes more than 170 hours and documents recording devices, transcripts and speaker-separated splits.

Use the released files and their actual subset definitions when estimating your training volume. Review pronunciation, vocabulary and recording fit for your application. Avoid assuming that a Mandarin read-speech corpus represents every accent, speaking style or production device you need.

TTS data: inspect text, speakers and recording files

13. LJ Speech: a single-reader starting point

LJ Speech 1.1 contains 13,100 clips from one reader, totalling about 24 hours, with a transcript for each clip. The publisher describes its public-domain status in the United States and notes that the original recordings were distributed as MP3 files.

Inspect the actual audio and text representation your pipeline will consume. A single-reader collection can help you explore a single-voice setup; a requirement for multiple distinct speakers calls for a different corpus or additional data. Check your model's preparation instructions before converting files.

14. VCTK: account for release-specific exceptions

VCTK 0.92 documents 110 English speakers with varied accents reading sentences. The release notes say transcripts are available for 109 speakers because p315's text was lost. They also identify microphone-specific technical issues for p280 and p315. Its linked license file specifies CC BY 4.0.

Check the actual microphone and text pairs before building a manifest. Use the documented speaker identifiers when designing an experiment, and retain a record of excluded files. The headline speaker count alone does not establish how many complete pairs your preprocessing will retain.

15. LibriTTS: sentence-level speech for synthesis

LibriTTS provides about 585 hours of multispeaker read English at 24 kHz. It derives from the original LibriSpeech materials and includes original and normalized text, with speech split at sentence boundaries. The publisher describes it as a TTS research corpus.

Compare both text forms with your model's expected input. If you also use LibriSpeech or other audiobook-derived data, check for shared source material when defining splits. A familiar dataset name does not replace checking the exact files and speaker coverage.

Turn the shortlist into a selection record

Write down the language or variety, speaking task, microphone conditions and output labels your application needs. For each candidate, identify the exact release, chosen training subset, transcript convention and any speaker or source exclusions. Keep development material for model choices and reserve separate material for final evaluation.

Then inspect representative audio and transcript pairs using the same preparation steps you plan to use later. Record missing text, unusable files and mismatched recording conditions. Estimate usable volume after that inspection rather than treating a corpus's headline hours as your project requirement.

For example, consider a fictional, untested English meeting-transcription prototype. Its brief asks for distant-microphone speech and speaker turns. AMI and CHiME-6 are candidates to inspect because they provide multichannel meeting or conversational recordings. The selection record would name the channels, annotation files and held-out meetings before comparing results. LibriSpeech could provide a separate read-speech comparison, while Speech Commands would address a different isolated-word task. This is a worked selection example, not a client result or a claim that any candidate will meet a target accuracy.

Check permissions and remaining coverage gaps

Save the release identifier, download source, license, attribution instructions and available provenance documentation. Review recordings, transcripts, software and pretrained models as separate components when their terms differ. Creative Commons' CC BY 4.0 deed explains that additional permissions, including privacy or publicity rights, may still be needed. A license label alone does not clear a particular deployment.

Public and commercial sources both need project-specific review. If a candidate lacks the language, recording conditions or labels in your brief, identify that gap before choosing more data. Our speech data provider comparison covers sourcing options and questions to ask a supplier.

For spontaneous conversation with transcripts, check the language and recording conditions in our dataset catalog. If you need speech recorded to a specific prompt list, device or environment, use the custom collection brief to describe those requirements. We assess the brief before confirming a collection plan.