ElevenLabs API

sonix exposes an ElevenLabs-shaped surface so existing ElevenLabs clients work by switching the base URL and key. These routes accept either Authorization: Bearer or the ElevenLabs xi-api-key header.

`POST /v1/text-to-speech/{voice_id}`

The path segment selects the voice. See Voices for how IDs resolve.

curl "$SONIX_URL/v1/text-to-speech/af_alloy?output_format=mp3_44100_128" \
  -H "xi-api-key: $SONIX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "text": "Hello from sonix.",
    "voice_settings": { "speed": 1.0 }
  }' --output hello.mp3

Field	Notes
`text`	The text to synthesize. Required.
`voice_settings.speed`	`0.25` to `4.0`. Other voice settings are accepted and ignored.
`model_id`	Accepted for compatibility; it does not select the sonix model.

Request bodies are lenient: unknown ElevenLabs fields (seed, previous_text, pronunciation dictionaries, request-id continuity) are accepted and ignored, so existing client payloads pass through unmodified.

Output formats

The output_format query parameter follows ElevenLabs naming and defaults to mp3_44100_128. The codec prefix maps to the response encoding: mp3_*, pcm_*, wav_* and opus_* are supported. Sample-rate suffixes are parsed best-effort; audio is not resampled. ulaw_* and alaw_* are rejected with a validation error.

On qwen3-tts-0.6b only pcm_* and wav_* are available. An omitted output_format returns wav_24000 (pcm_24000 on /stream) there. language_code defaults to English when omitted; it is never inferred from the text.

`POST /v1/text-to-speech/{voice_id}/stream`

Same contract as the non-streaming route, provided so ElevenLabs streaming clients work unchanged. The response is currently buffered: the full clip is synthesized and then returned. True chunked streaming is planned.