Home › Speech data collection
Speech data collection, sourced from real people who agreed to it
The hard part of a speech dataset is not the recording. It is finding the right humans, getting genuine consent, and being able to prove both when procurement asks. That is the part we are built around.
The sourcing problem nobody quotes for
Most speech data disappoints for the same reason: the speakers were whoever was available, not whoever the model needed. A platform with two million registered users still cannot give you forty Glaswegian speakers over sixty-five, or spontaneous Levantine Arabic from speakers who have never read a prompt script in their lives. It gives you what is in the pool and calls the gap an edge case.
Voices4AI exists to recruit against a brief rather than filter against a pool. When a project calls for accents, ages, dialects or speaking styles that are not sitting in a marketplace, we go and find those people, explain honestly what the recording is for, and pay them properly for it. That is slower than clicking a filter and it is the reason the datasets are usable.
How a project runs
Brief to delivered dataset
1. Specification
Languages, accents, speaker demographics, speaking style, recorded hours, technical format and delivery structure. If the profile is not realistically recruitable in your timeline, we say so before you commit budget rather than after.
2. Recruitment and consent
Speakers are recruited to the profile and take written, informed consent naming your organisation, the purpose, the territory and the licence term. Consent records are per contributor and retained as evidence, not as a checkbox.
3. Recording
Captured to broadcast standard - studio, partner facility or supervised remote depending on the project - with direction, monitoring and consistent technical spec across every contributor. Production is delivered through Flaunt Audio and the 11C Media studio group.
4. QC, annotation and delivery
Checked, cleaned and labelled to your schema, delivered in the structure your pipeline expects, with reject handling and retake policy agreed up front rather than argued about at the end.
Dataset types we collect
- Scripted read speech for TTS corpora - phonetically balanced, consistent, single or multi-speaker.
- Spontaneous and conversational speech for ASR and understanding - unscripted, overlapping, disfluent, which is how people actually talk.
- Multilingual and dialectal sets across the languages and accents we recruit.
- Voice identity capture for consent-cleared cloning, scoped and licensed explicitly.
- Targeted profiles - age bands, regional cohorts, second-language speakers, and other groups routinely missing from general corpora.
Provenance you can put in front of an auditor
Buyers are increasingly asked where their training audio came from, by procurement, by customers and by regulators. A dataset that cannot answer that is a liability sitting on a disk. Every project we run produces a consent and provenance trail alongside the audio, designed against EU AI Act Article 10 data governance and handled under UK and EU data protection law throughout.
Deeper studio production detail - facilities, capacity, technical process - sits with our parent group at 11c Media.
Commissioning questions
What is speech data collection?
The recruitment and recording of human speakers to produce audio for training or evaluating speech models - text-to-speech, speech recognition, speaker identification, wake-word detection and conversational systems. It is a sourcing and consent problem as much as an audio problem, which is why crowdsourced platforms so often deliver volume without usable coverage.
How is this different from a crowdsourcing platform?
Crowd platforms filter whoever is already registered. We recruit against the brief - the accent, the demographic, the language, the speaking style you actually specified - which is why the coverage looks different on a rare profile. Consent is collected per contributor, per project, with the client and purpose named.
How long does a collection project take?
It depends almost entirely on how rare the speaker profile is, not on the hours requested. Common languages and neutral accents move quickly. Dialectal, elderly, child, regional or second-language speakers take recruitment time, and we tell you that at quote stage rather than discovering it halfway through.
Can you deliver data our legal team will sign off?
That is the point of the business. Per-contributor written consent naming client, purpose, territory and licence term; documented provenance; no resale or repackaging; deletion routes in writing; and a data governance approach designed around EU AI Act Article 10 rather than retrofitted.
Tell us the profile, not just the hours
Send the speaker spec you actually need. We will tell you honestly whether it is recruitable and what it costs.