Audio and speech data: when a voice is a dataset and when it is a person
Accent-diverse, well-transcribed English speech is useful for recognition models. A voice can also identify a person, and synthetic voice use carries its own risks.
When a voice is data and when it is a person
- Recording Source, notice and distribution terms.
- Speaker Consent, voiceprint and cloning limits.
- Description Hours, speakers, transcripts, sample rate.
Common sources and their main issue
| Source | Main question |
|---|---|
| Contact-center recordings | Callers heard 'recorded for quality and training' — did that mean AI model training for third parties? |
| Podcast and radio archives | Guest, host and music rights; distribution-only agreements |
| Purpose-recorded speech corpora | Cleanest source when consent names model training |
| Oral history collections | Donor agreements, estates, institutional terms |
Voice cloning is a different category
If a dataset could be used to imitate a specific speaker, generic research consent is not enough. Expect to need explicit, informed consent and contractual bans on impersonation. State biometric laws may treat voiceprints as biometric identifiers.
Describe in an inquiry
- Total hours and speaker count.
- Transcript availability and who produced them.
- Accent, age and recording-environment mix.
- Sample rate and channel layout.
How this affects value
Consent that names model training—and excludes impersonation—makes a speech corpus far easier to license.
See what shapes a dataset's value