Training infrastructure for a text-to-speech generation system

Design the data and training workflow for a text-to-speech service. Use the 2024 Fish Speech architecture as a reference, distinguish audio-code prediction from waveform decoding, and make duration, token count, and throughput assumptions explicit.

Text-to-speech (TTS) systems turn written text into spoken audio. The words should be intelligible, while pronunciation, pauses, emphasis, and voice characteristics should fit the intended use. A waveform can sound natural and still read a name or number incorrectly.

Tacotron 2, XTTS, and Fish Speech illustrate different approaches to speech synthesis. Applications include voiceovers, accessibility tools, and spoken interfaces. A translation application also needs a translation stage; TTS alone speaks the supplied text. This case does not claim that a particular public assistant uses one of these architectures.

The example below illustrates the service boundary: text and supported voice settings enter, and playable audio leaves. Define the supported behavior before selecting the model.