Voices

GET /v2/voices

Lists the voices of the active speech model in the paginated ElevenLabs shape:

curl "$SONIX_URL/v2/voices" -H "xi-api-key: $SONIX_API_KEY"

{
  "voices": [
    {
      "voice_id": "af_alloy",
      "name": "Alloy",
      "category": "premade",
      "labels": { "gender": "female", "language": "en" }
    }
  ],
  "has_more": false,
  "total_count": 1,
  "next_page_token": null
}

The legacy GET /v1/voices returns the same list as { "voices": [...] }.

The catalog reflects what the deployment you target actually serves; list voices first to discover what your endpoint offers.

Voice resolution

Wherever a voice is named (the voice field on /v1/audio/speech or the {voice_id} path segment on the ElevenLabs surface), sonix resolves it in order:

1. An exact match against the model's voice catalog (af_alloy), then a case-insensitive match, then a catalog display-name match (Alloy).
2. The literal default selects the model's default voice.
3. An OpenAI-style name (alloy) maps to the catalog ID with the same suffix (af_alloy) when one exists.

So OpenAI's stock voice names work out of the box on models whose catalogs follow that convention. Exact catalog IDs always win.

Default voices

Every voice below reads the same line, so what you hear between two rows is the voice and nothing else. Press ↑ and ↓ to step through a catalog, playing each voice as it lands.

kokoro-82m

Style-vector voices covering American and British English.

OpenAI's alloy, echo, onyx and nova resolve by suffix. fable, shimmer, ash, ballad, coral, sage and verse have no counterpart here; they return a validation error rather than silently remapping to a different voice.

kitten-tts-nano-0.8, kitten-tts-micro-0.8 and kitten-tts-mini-0.8

All three sizes share the same eight voice ids and display names, but they are different models and do not sound the same — mini reads the same line more slowly than nano, with micro between them. Compare the tables before picking a size.

qwen3-tts-0.6b

Nine preset speakers baked into the model. Voice cloning is not available. Address each speaker in its native language for best results — which is why the Chinese, Japanese and Korean speakers below are auditioned in their own language rather than in English.

pocket-tts

One predefined voice baked into the model. This model is English only. Voice cloning is not available: the deployed bundle is upstream's no-cloning release, so Alba is the entire voice surface. The voice is built from Alba MacKenna's recordings, released under CC BY 4.0.

neutts-nano-q4

Voices are cloned from reference recordings. This model is English only. Upstream ships multilingual NeuTTS as separate per-language checkpoints; the one served here is the English base, so send it English text.

Upstream publishes nine reference speakers; the six English ones are offered here. The other three are cloned from German, French and Spanish recordings, and they are omitted rather than sold as accented English: NeuTTS phonemizes a voice's reference transcript with the same English-only phonemizer it uses for your text, so a non-English reference degrades the clone no matter what you ask it to say.