Voice data for AI teams
Speech data recorded by professional voice actors who gave written consent for this specific use. Cast to your languages, accents and spec.
Why this data holds up
Most speech data was collected without anyone’s permission. Ours is recorded to a framework that is published in our terms and applied to every dataset, with no exceptions.
Informed written consent from the specific actor for the specific dataset and use. Consent given in general terms does not count.
The agreement names the use, the term and the territory. Paid advertising and voice cloning are excluded unless separately licensed.
Actors are paid for the life of any model that uses their voice. Never a flat buyout, and never a session fee standing in for a licence.
Ownership of the voice identity stays with the actor. They can decline a renewal, and they can decline a project without giving a reason.
What you get
Typical builds
Many speakers, short scripted utterances, with a wide accent spread, controlled noise conditions and the phonetic coverage your spec calls for.
Fewer speakers, the same content across emotions, with labelled intensity levels that are distinguishable in QC.
Small, controlled, held out of any training pool, with the composition you specify and documented speaker coverage.
One actor, a named product and term, licensed separately from any dataset work, with the actor signing per renewal.
Every actor here opted in to data work. Filter by language, gender, voice age and pitch, play a sample, and open a row for more demos. Actors are identified by Dataset ID until a project is agreed. Opting in to this list is not consent to a specific use: consent is signed per dataset, and an actor can decline any project.
205 voices found
19 more samples
11 more samples
8 more samples
12 more samples
12 more samples
How it works
You tell us the languages, accents, labels and accepted minutes. We turn that into a script, a recording spec and a sample for you to approve.
A paid pilot proves the spec on real audio before the full build. It is credited in full against the project that follows.
We cast from the opted-in roster. Each actor signs for this dataset and this use, and gets a consent ID before recording starts.
Directed remotely, in the actor’s own studio, with a native speaker on pronunciation and locale where the language needs one.
Technical, linguistic, performance, label and consent checks. Clips that fail are re-recorded at our cost, not billed to you.
WAV plus a metadata CSV and checksums, delivered to named recipients only, with the permitted use written into the readme.
Straight answers
What every agreement covers, and what we will not do at any price. Worth reading before you brief us, so neither of us wastes a call.
One agreement per actor, per dataset, so there is a separate signature behind every voice in your set. The agreement itself goes out after the first call, once we know what we are building.
These are declined whatever the budget. If a project needs one of them, it is better that we both find out now.
An off-the-shelf corpus usually comes with a licence and no way to see what each speaker agreed to. Here, every clip carries a consent ID that resolves to a signed agreement naming the dataset, the use, the term and the territory. The data is also recorded to your spec rather than selected from whatever exists.
Price depends on the number of actors, accepted minutes per actor, the language tier and the QC level. Accepted minutes are minutes that pass QC, and rejected takes are re-recorded at our cost. Tell us the shape of what you need and you get a figure, usually the same week.
Yes. A paid pilot gives you real audio, real metadata and a real QC report against your spec, and it is credited in full against the project that follows.
Sometimes, and never by default. It is a separate licence from any dataset work, it names the product, the term and the territory, compensation continues for as long as the voice is in use, and the actor can decline. Nothing in a standard recording grants cloning or synthesis rights.
Their clips come out of the dataset and out of any training pool, and we confirm it in writing. How models already trained are handled is written into the agreement before anyone signs, so it is settled in advance rather than argued about later.
Our roster covers 80+ languages. The voices listed above are the actors who have opted in to data work so far, and that list grows. If the language you need is not showing, ask, because casting for a specific dataset is part of the service.
WAV at the sample rate, bit depth and channel count in your spec, segmented to your clip length, with a metadata CSV, a readme stating the permitted use, and checksums. Access is limited to recipients you name.
Languages, accents, accepted minutes and the labels you need. We come back with a shape, a timeline and a price.
