Open speech datasets provide or reference recordings and, depending on the release, transcripts, speaker labels or other annotations. LibriSpeech, Common Voice and VoxCeleb serve different research tasks. Start with the exact release and your model training or evaluation job: a corpus useful for one benchmark may lack the labels, recording conditions or permissions your project needs.
This guide compares documented corpus designs and shows how to turn a possible limitation into a checkable sourcing decision. An open release that meets the required task, coverage and permitted uses can remain the right choice. Another source or additional annotation is worth assessing when you identify a specific unresolved gap.
For a procurement decision, compare an open release with a specific listed asset. Browse Spirelight's conversational dataset catalogue and request a representative sample; confirm availability, coverage, permissions and written terms before selection.
Compare open speech datasets by task
Choose the dataset for the work you need to test or train, then inspect the exact release. This comparison is a starting point for model training and evaluation, not a finding that any source is suitable for your project or permitted use.
| Dataset | Starting task | First check |
|---|---|---|
| LibriSpeech | Read English speech recognition | Benchmark splits and fit with your recording conditions |
| Common Voice | Scripted or prompted speech, depending on series | Version, validated hours, transcripts and usage information |
| VoxCeleb | Speaker recognition research | Labels, evaluation design and source permissions |
| AudioSet | General sound-event recognition | Which labels and source media you can actually use |
| GigaSpeech | English recognition across mixed source styles | Subset, transcripts, splits and audio permissions |
Release notes and primary sources
- LibriSpeech: About 1,000 hours of 16 kHz read English from LibriVox audiobooks. OpenSLR documents CC BY 4.0 and separate train, development and test splits, including clean and other subsets. Preserve the prescribed splits for benchmark comparisons; a score on them does not establish performance on your recordings.
- Common Voice: Mozilla has scripted and spontaneous series. Its 18 September 2026 announcement names Scripted 27.0 and Spontaneous 5.0 as dated examples. The English Scripted 27 page separates recorded from validated hours. The English Spontaneous 5 page describes prompted responses, some without transcripts. Both show CC0 labels and separate usage information. Check the exact release, access conditions and speaker-privacy questions.
- VoxCeleb: Interview-video segments for speaker-recognition research. Check the official documentation for the version, speaker labels and evaluation design. Its privacy notice describes the university's processing, not a licence or privacy exemption for your organisation. Confirm source and permission records for your intended use; interview provenance alone does not settle individual consent.
- AudioSet: Google's overview describes a 632-class sound ontology and about two million labelled ten-second video segments. Its download page describes 527 classes in the annotated sets, plus segment IDs, labels and features. The two class counts refer to different things. Rights in released annotations or the ontology do not establish rights in each underlying video.
- GigaSpeech: The 2021 v1 documentation describes 10,000 hours of transcribed English from audiobooks, podcasts and online video, including read and spontaneous sources. Inspect the subset, source conditions, transcript preparation and splits. A code-licence badge alone does not establish permissions for the audio.
Other task-specific options
- MLCommons' People's Speech describes a large transcribed English release with stated attribution and share-alike components.
- LJ Speech documents about 24 hours of read speech from one speaker, paired with text; it is not a multi-speaker conversation source.
- The Google FLEURS card describes parallel sentence-based evaluation data across 102 languages.
- The TED-LIUM 3 paper documents TED-talk-derived English recognition data. Check a current release's access and terms separately; a historical paper does not certify them.
What open datasets are good for
Match a documented release to a specific job: reproducing a published recognition benchmark, testing a speaker-recognition system, exploring synthesis with paired audio and text, or inspecting sound-event labels. Record the subset, version, preprocessing and evaluation protocol so another person can understand the result. A comparison is only meaningful within the conditions actually tested.
Open material may also support prototyping, training or adaptation when its contents and terms fit the proposed method. Do not assume one required sequence from open pretraining to paid fine-tuning. An existing model, a small evaluation set, extra labels or a different release may address the immediate need. Define the gap before deciding how many additional hours to source.
Where open datasets stop being enough
- Speech style and capture: distinguish read, prompted and spontaneous speech from microphone distance, channel, device and noise. LibriSpeech's read audiobook design and Common Voice's separate series illustrate why the corpus name alone is insufficient. For a conversational task, inspect actual turns, interruptions and any required overlap labels; the conversational speech data guide covers that brief.
- Labels and validation: total audio duration is not necessarily the amount of validated, transcribed or speaker-labelled material available for your task. Verify the release's definitions and inspect a permitted sample. Missing turn boundaries, for example, may call for additional annotation rather than new recordings.
- Coverage: compare language, locale, speakers and recording conditions with your intended evaluation. A language name or overall hour count does not establish coverage of a particular dialect or device. Avoid treating contributor demographics as representative of your users without evidence; request appropriate records where available and justified.
- Source and permissions: check the exact release's licence, attribution or share-alike duties, source and platform permissions, privacy information and applicable lawful basis. Keep unresolved rights questions separate from technical fit. The speech data licensing guide lists questions to investigate; it does not replace the actual source records or project-specific review.
- Evaluation independence: check what is known about the model version's training data and the proposed split. If prior exposure is unknown, record that limitation and consider an appropriately held-out evaluation source. A benchmark score remains evidence about that test; it does not establish performance across every deployment condition.
Turn a limitation into a sourcing decision
The following examples are hypothetical planning decisions, not tests we ran, client results or claims about a supplier's available stock. Each starts with evidence you can request or inspect rather than assuming all open or commercial data has the same strengths.
| Observed evidence | Decision to resolve | Practical next step |
|---|---|---|
| A release lists total audio and a smaller validated transcript subset. | Which subset meets the supervised recognition brief? | Scope the qualifying subset first; assess additional transcription or another documented source only for the remainder. |
| An authorised conversation sample has transcripts but no speaker-turn labels. | Can the existing recordings support the diarization task? | Check speakers, channels and overlap, then price the required annotation and acceptance process before commissioning replacement audio. |
| The files and labels meet the technical checks, but source permissions remain unresolved. | Can this specific use proceed? | Keep the rights decision open for the applicable review. A successful model test or paid invoice does not close it. |
Record the exact release, intended task, required coverage, label definitions, test protocol, source records and unresolved questions in the speech dataset specification worksheet. Keep the original source identifiers and a record of transformations so that the candidate you assess is the one your team proposes to use.
When to license or commission data instead
If an open release meets the required coverage, permitted uses and evaluation needs, use the supported option. If a material gap remains, compare preparing or annotating that source, selecting another documented asset, and a feasible new collection. Consider the whole proposed scope and acceptance work; neither a free download nor a commercial label establishes the final implementation effort or suitability.
Spirelight's catalogue has 60 conversational speech dataset listings with recording details and sample-request options. Confirm current availability, sample coverage, permitted uses and written commercial terms for the specific asset. The buying AI training data guide covers comparing a listed dataset with a custom recording scope. Custom collection needs a feasibility assessment and agreed recruitment, recording, annotation, responsibilities, schedule and acceptance; a missing open-data feature is not proof that any supplier can deliver it.