What is TTS?

TTS wants the opposite of what ASR wants: clean, stable recordings with a consistent voice. The same sentence is best recorded many times by the same speaker in the same state.

That is why audiobook-style narration is worth so much in TTS training: long sessions, one consistent voice, clear pronunciation.

Varieties of TTS training data

TTS data divides by how many voices it holds. A single-speaker set goes deep on one voice: many hours, one person, consistent conditions, enough material to cover the sounds and the sentence rhythms of the language. A multi-speaker set goes wide instead, many voices and usually shallower on each, for systems that must speak in a voice they were not built around.

The second division is how much expression is in it. Neutral narration is the base: a speaker reading prepared text in a steady tone. Expressive material adds labels for emotion, speaking style or intensity, and it is harder to produce because the label is itself a judgement. Listeners disagree on whether a line sounds annoyed or merely firm. Whispered and shouted speech are separate categories again.

The third is the pairing between audio and text. TTS training needs the exact words that were spoken, written the way they were read. A sentence containing a date, an amount of money or an abbreviation has to be written in the expanded form the speaker actually said, or the model learns to read the symbols rather than the words. That normalisation belongs in the specification.

What to specify for a TTS voice

Start with the voice itself. Language, regional accent, gender, age range and register, whether narration, customer service or conversation, all change the pool of speakers who qualify and therefore the price. If the brief says native speaker, define it: someone who grew up speaking the language in the region, or someone fluent who learned it later. The two sound different.

Then the recording conditions, because TTS is unforgiving about them. The audio is the product, so room tone, microphone distance and session noise are audible in the output rather than merely dragging down an accuracy number. TTS is usually wanted at a higher sample rate than speech recognition, since the result has to sound good to a listener. One voice should be recorded across sessions on the same equipment.

Rights need spelling out here more than anywhere else in speech data. Transcribing a recording and synthesising a voice from it are different uses, and consent for the first does not cover the second. The brief should also say whether the buyer receives the raw takes with aligned text, and whether the supplier will re-record gaps found after delivery.