xAI

xAI tuning is passed in synthesizer.options.

1 {
2   "verb": "say",
3   "text": "Your balance is $42.50.",
4   "synthesizer": {
5     "vendor": "xai",
6     "language": "en",
7     "voice": "eve",
8     "options": {
9       "speed": 1.0,
10       "optimize_streaming_latency": 0,
11       "text_normalization": false
12     }
13   }
14 }

You can also set these in the portal under Extra Options on the speech credential, in which case they apply to every request that uses that credential. Settings supplied on the verb take precedence.

An xAI credential can be enabled for both TTS and STT. Extra Options applies to TTS only — speech recognition settings are configured per-verb via recognizer.xaiOptions, not on the credential.

Options

speed

number

Speech speed multiplier, 0.7 – 1.5. Defaults to 1.0. Values below 1.0 slow the voice down, above 1.0 speed it up.

optimize_streaming_latency

number

0 (default) or 1. 1 shrinks the first audio chunk to lower time-to-first-audio, with a minor quality tradeoff.

text_normalization

boolean

Convert written-form text — numbers, abbreviations, symbols — into spoken form before synthesis, so $42.50 is read as “forty-two dollars and fifty cents”. Defaults to false.

xAI is one of the vendors jambonz uses word-level timestamps with, to trim already-spoken audio on barge-in. Normalization rewrites the text that is actually spoken, which makes the returned timestamps disagree with the text jambonz sent. When timestamps are in use, text_normalization is therefore suppressed — alignment wins. If you need normalized readings of numbers or currency in an agent, write them out in the text you send rather than relying on this option.

Endpoint override

xAI credentials have no API base URL field in the portal. If you need to point at a non-default endpoint — a proxy, for example — set api_uri in Extra Options alongside the tuning fields:

1 {
2   "api_uri": "my-proxy.example.com",
3   "speed": 1.0
4 }

Leave it out to use xAI’s standard endpoint (api.x.ai).

Streaming

xAI supports TTS streaming, so it can be driven token-by-token from an LLM using tts:tokens. Word-level timestamps are returned, so alignment events are available — see the warning above for how that interacts with text_normalization.