Text-to-speech (TTS)
Also known as：TTS · speech synthesis · voice synthesis
Text-to-speech (TTS) is the technology that converts written text into audio. In live translation it reads the translated text aloud, and because its cost scales with text length it is one of the largest contributors to end-to-end latency.
Where it sits in the pipeline
TTS is the fourth stage, immediately after machine translation. Its output goes into the virtual audio cable and from there through OBS to the audience.
What makes it unusual is that TTS costs both generation time and playback time. However fast generation gets, a sentence of a dozen or so English words still takes several seconds to say, and that part depends on sentence length rather than on the engine. There is only one way to shorten it: write shorter sentences.
Parameters that change the experience
- Rate. Google Cloud Text-to-Speech’s
speakingRateaccepts 0.25 to 2.0, where 1.0 is the voice’s native speed. Nudging it up to around 1.1 shortens playback proportionally; how much naturalness you lose depends on the voice and the language, so listen before committing. - Voice. Stock and cloned voices differ not just in naturalness but in latency.
- Pause handling. Good synthesis places sensible pauses at punctuation. If the translation loses its punctuation, delivery runs together and intelligibility drops sharply.
- Numbers and units. Whether “30ml” is read as “thirty millilitres” or “three zero m l” varies by engine. Prices and specifications appear constantly in live selling, so this is worth testing on its own.
Streaming synthesis
Like recognition, synthesis comes in streaming and non-streaming forms. Streaming synthesis begins playback before the full sentence has been generated, which noticeably reduces perceived latency. If your stack supports it, turn it on.