Speech and audio data for AI training
Recorded speech and other audio with transcripts, timestamps, speaker labels and metadata on language, accent and recording conditions.
- Measured in
- Hours of audio, number of speakers, languages and accents.
- Who buys
- Speech recognition vendors, voice AI companies, labs training audio and multimodal models.
- Where it comes from
- Podcast networks, broadcasters, call centers, transcription services, voice-over libraries and research institutions.
- Rights to clear
- Speaker consent and release forms, guest agreements, music and sound effect licenses, and union agreements for performers.
- Privacy
- Voices can identify people. Illinois BIPA covers voiceprints, so collection consent and purpose matter.
- Pricing signal
- Defined.ai cited $100 to $300 per hour for longer video and film (Reuters, 2024). Speech is priced by hour, quality and transcript accuracy.
- Speech datasets train speech recognition, voice assistants and audio language models, and buyers value accent, language and acoustic variety.
- Transcripts aligned to the audio make a recording far more useful than audio alone.
- Voice is biometric information in some US states, so consent records matter as much as the recordings.
- Podcast and broadcast archives are licensable when the holder owns the rights to every voice and music bed in them.
What buyers look for
- Variety. Accents, ages, genders, languages and recording environments (studio, phone, street, car).
- Transcripts. Accurate, time-aligned transcripts with speaker turns. Human-verified transcripts are worth more than raw machine output.
- Natural speech. Spontaneous conversation teaches more than read scripts, which is why podcasts and call recordings are in demand.
- Clean rights. A consent or release record for every speaker.
Consent and biometric law
A recording of a voice can be used to identify a person. Illinois's Biometric Information Privacy Act regulates voiceprints and requires informed written consent for collection (Illinois General Assembly). Not every speech dataset involves voiceprints, but buyers will ask how consent was collected and what speakers were told. Performer unions have also negotiated terms on digital replicas of voices, which is why many licenses forbid voice cloning.
Podcast and broadcast archives
A podcast network may own the episodes but not every voice in them. Guest release forms, co-host agreements and music licenses all affect what can be licensed. A common structure licenses host-only segments first, then adds guest audio where releases allow.
For call recordings
Call center audio has its own rules on recording consent and customer privacy. See call center recordings.
Questions
How much is an hour of speech data worth?
It varies widely with language, transcript quality and rights. Reuters reported one vendor paying $100 to $300 per hour for longer video and film footage; speech without video is negotiated separately. Rare languages and accents command more.
Do I need consent from every speaker?
Buyers will expect a clear legal basis for every voice, which usually means release forms or contracts. Where consent is missing, that audio is normally excluded or de-identified.
Can licensed voices be cloned?
Most licenses forbid creating synthetic replicas of identifiable speakers. Make that restriction explicit in your agreement.
Related guides
Related data types
Sources
Need speech and audio data, or hold it?
Tell us what you need or what you hold. We check rights and privacy first, then introduce vetted counterparties. You license directly with each other.
- We reply to every enquiry within two working days.
- No upfront fees for sellers
- We never hold your raw data