Rime Labs

Rime Labs tuning is passed in synthesizer.options. Which fields apply depends on the model — Rime’s model families expose quite different controls.

1 {
2   "verb": "say",
3   "text": "Thanks for calling, how can I help you today?",
4   "synthesizer": {
5     "vendor": "rimelabs",
6     "language": "eng",
7     "voice": "luna",
8     "options": {
9       "speedAlpha": 1.0,
10       "reduceLatency": true
11     }
12   }
13 }

mist, mistv2 and v1

speedAlpha

number

Speaking rate. Defaults to 1.0. Values above 1.0 are slower and below are faster — the opposite of a multiplier, so 0.9 speeds the voice up.

reduceLatency

boolean

Defaults to true. Reduces time-to-first-audio at a small cost in prosody.

pauseBetweenBrackets

boolean

Treat bracketed numbers in the text as pause durations in milliseconds — <200> inserts a 200 ms pause.

phonemizeBetweenBrackets

boolean

Treat bracketed text as explicit phonemes rather than words, for controlling pronunciation of names and jargon.

inlineSpeedAlpha

string

Per-segment speed values, letting one utterance change pace partway through.

arcana

Arcana is a sampling-based model and takes the usual generation controls instead:

temperature

number

Sampling temperature. Higher values give more varied delivery.

top_p

number

Nucleus sampling cutoff.

repetition_penalty

number

Discourages repeated audio patterns.

max_tokens

number

Caps the audio generated for one request.

coda

timeScaleFactor

number

Playback speed — above 1.0 slows down, below speeds up. Applies to the non-streaming HTTP path only, so it affects say but not token streaming.

segment

string

How text is chunked over the streaming websocket: immediate, never, or bySentence. bySentence generally gives the best prosody; immediate the lowest latency.

These three groups are not interchangeable. Sending speedAlpha to arcana, or temperature to mistv2, has no effect rather than raising an error — check which model the credential uses before concluding a value is wrong.

Streaming

Rime Labs supports TTS streaming with word-level alignment and a native clear for barge-in. Its flush ends the vendor connection, so jambonz redials per turn.