Definition
Text-to-Speech converts written text into natural-sounding speech. A TTS dataset therefore consists of recordings of a chosen voice reading known text, captured under tightly controlled conditions and aligned to that text.
Consistency beats volume
Because a TTS model learns the character of one voice, variation in microphone, room, distance, level or delivery becomes noise the model has to fight. A smaller, rigorously consistent corpus often outperforms a larger, sloppier one.
- ▸Same microphone, chain and room across every session
- ▸Stable levels and mic distance
- ▸Consistent speaking style and pace
- ▸Clean starts and ends, no background events
- ▸Script-aligned transcripts with prompt IDs
Script design
Scripts should cover the phonetic and prosodic range the voice must produce, plus domain vocabulary: numbers, dates, names, addresses, product terms and loanwords. For Thai, tone patterns and transliterated English need deliberate coverage.
Consent and rights come first
Synthetic voice work needs explicit permissions from the speaker covering AI training and the commercial scope you intend. Agree this before the first session — it cannot be widened later.
Related services