Generate Speech (Kokoro)

How it fits

Typical backends

Caveats

• I/O contractOutput is always 24 kHz mono; there is no rate knob, so consumers that need another rate must resample. Long inputs are segmented internally and the segments are concatenated, so one call yields the whole utterance as a single frame.
• I/O contractAn empty input String, or text that produces no audio segments, returns an empty mono AudioFrame without raising — silent gaps surface as zero-length frames.
• Hard constraintvoice must name a real Kokoro preset; an unknown name is not caught at startup and fails inside the pipeline on the first call that uses it.
• Accuracylang_code selects the Kokoro language frontend; pairing it with text in a different language usually yields unintelligible audio rather than an error.
• FallbackRequesting CUDA through device on a host with no visible GPU continues on CPU at much slower speeds.
• State lifetimedevice and lang_code bind the pipeline once at startup, whereas voice is read fresh on every call and can be switched per request without redeploying.