Generate Speech (Voxtral)

How it fits

Typical backends

Caveats

• Hard constraintThis component is a thin client over a colocated vLLM service that must be serving the Voxtral model; the service is declared as a dependency so it is deployed alongside, and the component container itself needs no GPU.
• Hard constraintThe vLLM host and port are read from the environment at startup rather than exposed as knobs; if they are unset the client cannot reach the service.
• Hard constraintvoice must be one of the baked-in presets and model_name must be the Voxtral checkpoint the service has loaded; an unrecognised voice or model is rejected by the service as an HTTP error rather than caught locally.
• Parameter interactionresponse_format selects the wire encoding requested from the service; the bytes are decoded back to samples and the source sample rate is preserved verbatim with no resampling.
• LatencyThe request to the service uses a fixed two-minute timeout; very long inputs that exceed it fail with a network error rather than returning partial audio.
• I/O contractAn empty or whitespace-only input String short-circuits to an empty AudioFrame without contacting the service.
• CompatibilityConfirm the Voxtral model licence permits your intended use; the default checkpoint carries a non-commercial licence.