Generate Speech (Chatterbox)

How it fits

Typical backends

Caveats

• I/O contractThe output sample rate is whatever the loaded checkpoint reports and is not surfaced as a knob; downstream stages that assume a specific rate must resample.
• I/O contractA whitespace-only or empty utterance short-circuits to an empty mono AudioFrame at the model rate without ever running the decoder, so silent gaps in a text stream emit zero-length frames rather than silence.
• Hard constraintmodel_name must point at a Chatterbox-Turbo-family checkpoint; the Turbo loader expects that weight layout, and unrelated repos load but decode to noise.
• Parameter interactionPushing temperature toward zero collapses decoding to greedy and flattens prosody, while higher values widen variation at the cost of stability; top_p and top_k clip the candidate set on top of that and repetition_penalty discourages token loops.
• FallbackRequesting a CUDA device through device on a host with no visible GPU continues on CPU at far slower speeds instead of failing.
• State lifetimeEvery knob, including norm_loudness, is read once when the decoder is built at startup; there is no per-call re-read, so any change needs a redeploy.