Home › TTS training data
TTS training data recorded for consistency, licensed for keeps
A synthetic voice inherits every inconsistency in its corpus. Room changes, mic drift, energy shifts and processing all end up as artefacts you cannot fix in the model. Here is how we build corpora that do not fight you later.
What we deliver
- Single-speaker TTS corpora for a branded or product voice, recorded to unvarying spec across the whole capture.
- Multi-speaker corpora for general models, with controlled variation across speakers and controlled consistency within each.
- Expressive and emotional sets where the voice needs range rather than one neutral register.
- Phonetically balanced prompt design for the target language, written or adapted as part of the project.
- Raw unprocessed WAV at 48kHz/24-bit or above, with per-utterance metadata and transcripts to your schema.
Why we refuse to process the audio
Noise reduction, compression, gating and EQ all discard information and add artefacts that a model will learn as if they were features of the voice. A commercially “finished” voiceover file is the wrong deliverable for TTS training. We capture clean at source so nothing has to be repaired afterwards, and we deliver raw so your pipeline decides what happens next.
The licence is the risk
A TTS corpus is the highest-stakes licence in speech data, because the output is a voice a real person can recognise as theirs. That is the scenario contributors are most anxious about and the one regulators are most interested in.
So every TTS engagement runs on written consent that names your organisation, the purpose, the territory and the term. Contributors know a synthetic voice will be built. They are paid for that specifically, not slipped into it under a general dataset agreement. It costs more than the alternative and it is the only version that survives a due-diligence call. Our full position is on ethics and consent, and the talent side is set out in what we tell contributors about cloning.
Studio capacity, facilities and production process are detailed by our parent group at 11c Media.
TTS corpus questions
What makes a good TTS training corpus?
Consistency above almost everything. One speaker, one room, one microphone position, one distance, one energy level, across every hour recorded. Phonetic coverage of the target language, clean prosody, and raw unprocessed audio at 48kHz/24-bit or higher. Variation that would be invisible in a voiceover session becomes audible artefacts in a synthesised voice.
How many hours do we need?
It depends on the architecture and how much you are fine-tuning versus training from scratch, so anyone quoting a universal number is guessing. What we can say is that corpus quality and consistency change the required hours far more than raw volume does - clean data reduces the hours you need.
Can we license the voice commercially and permanently?
Commercially, yes, with the licence scoped and priced accordingly. Permanently, we would rather structure a long defined term with renewal than a perpetual buyout, because perpetual grants are exactly the terms that make contributors refuse and make future audits uncomfortable.
Do you provide the scripts?
We can write and phonetically balance the prompt set for the target language, work from your scripts, or adapt yours where coverage gaps show up. Script design is part of the dataset design, not an afterthought.
Tell us what your model needs to sound like
Send the spec - language, speaker profile, hours, licence intent - and we will come back with an approach.