xAGI LABS / DATA PROGRAMS
Speech and audio training data
Scope speech data against language coverage, recording conditions and the model task.
Start with the acoustic and language gap
ASR evaluation, conversational voice systems and speech generation can require different collection and annotation choices. Specify the language and dialect, speaking style, number of speakers, utterance duration and acoustic conditions. Separate the coverage you need for training from the examples reserved for evaluation. Quantity alone does not describe whether a speech corpus suits the model.
Define the recording and transcript checks
Agree the audio format, channel layout, sample rate, clipping and background-noise criteria, transcript conventions and speaker metadata. Document how consent and recording permissions will be assessed. Define acceptable transcription error, annotation review and a rejection workflow before expanding a collection. Speaker-level separation and duplicate checks help keep evaluation distinct from training.
Use a representative pilot
A pilot should include the hard conditions in the brief, not only the cleanest recordings. Review intelligibility, coverage and the metadata required by the training pipeline. Our Haryanvi TTS project documents research workflows; it is not a ready-made commercial dataset offer. Source availability, participant permissions and delivery timing are assessed per project.
An example speech datasheet
Illustrative planning template, version 1.0 · 17 October 2026. These examples are not a dataset listing, measured project results or a delivery commitment.
| Field | What the project must establish |
|---|---|
| Task and split | ASR evaluation; speaker-disjoint holdout; list intended languages and dialects. |
| Capture and coverage | Specify sample rate, channels, environment, speaking style and coverage quotas in the brief. |
| Provenance and rights | Record source authority, collection context, consent scope and permitted training/evaluation uses; verify supporting evidence. |
| Quality and rejection | Agree clipping/noise limits, transcript conventions, duplicate checks and rejection rules before the pilot. |
| Delivery and version | Manifest, audio/transcript identifiers, schema version, checksums and documented exclusions; no contributor identities in public examples. |
Check a representative pilot against each agreed gate. Quarantine unmatched files; report rejected items and review decisions. Separate consent evidence from public manifests and keep speaker identifiers pseudonymous.
Download the speech datasheet template (Markdown) →
Prepared by the xAGI Labs team for project scoping. A project-specific domain reviewer and acceptance criteria are agreed before collection or labeling.
Start with a useful conversation.
Tell us the model task or source you want to explore. We establish feasibility, quality and rights before discussing delivery.
Book a scoping call ↗