Deploying a text-to-speech generation model

Design a text-to-speech serving path from validated text to playable audio. Separate model interfaces from optional linguistic services, and estimate capacity using measured generation time and requested audio duration.

The service accepts text and supported voice settings and returns speech. Use the preceding lesson’s Fish Speech discussion as one architectural reference, while recognizing that the explicit phoneme and mel-spectrogram pipeline shown here is an alternative decomposition. It is not a literal list of separate services inside Fish Speech.

First define the workload and its storage assumptions.

Storage estimation

Assume 100 million daily active users making ten requests each per day, and separately assume 100 million stored profiles. Each request produces ten seconds of audio. Use decimal byte units.

• Raw weights: For the hypothetical two-billion-parameter planning model, FP16 values occupy four decimal GB. This is not a count inferred from the task being speech rather than images, and it excludes separate decoder components and runtime state.

• Profiles: Assume 10 KB per stored account: