Class TranscriptionService

The resolved checkpoint id; only meaningful after resolveModel() returned one.

Weights are loaded right now.

Stop admission, cancel pending work, join every session, take, creation and load in flight, then release the speech gate. The host releases the weights. Idempotent (later calls return the same promise); with nothing in flight the release happens synchronously.

Lease the model, loading it if it is not resident. The host's failures become HTTP-shaped ones.

called by the session when it closes; its lease is already back with the host.

The checkpoint this service leases, resolved once (as in main, a model downloaded later needs a restart); null when the host has none.

window work runs FIFO with requests; the decode call takes the host's lock.

Release the model now (GPU memory returns; the next request reloads). False while a take or session is using it, or when it is not loaded.

The Silero gate, loaded on first use; a gate arriving after close is released.

Fit vocabulary terms into the prompt budget, ", "-joined in order, whole terms only.