izwi tts --speaker, the Text to Speech page, or
the voice field on /v1/audio/speech (or speaker on /v1/text-to-speech).
Saved and reference voices use --saved-voice-id, reference_audio, or the
Voice Studio saved voice library instead of these built-in speaker IDs.
Model Families
Qwen3 CustomVoice
Example:
Kokoro-82M
Kokoro voice IDs use a two-letter prefix:
Available IDs:
Voxtral TTS
LFM2.5 Audio
LFM3 Audio
LFM3-Audio-2B-GGUF and LFM3-Audio-2B-Q8_0-GGUF have two voices. Voice
names are not case-sensitive, and a request without one uses female.
Either voice speaks any of the 19 locales that
language (--language in the
CLI) names: en-US (American English, the default), en-GB (British
English), ar, da, de, el, fi, fr, he, it, ja, ko, nl,
pl, pt, es, sv, tr and zh. Any other language returns 400.
Example:
Chatterbox
See Also
Maya1
Maya1-3B has no recorded voices: each built-in voice is a description in
the style the model was trained on, and the model invents a voice to match
it. A request’s own description (instructions, or --voice-description in
the CLI) replaces the voice. The same description and seed give the same
audio for the same text; across different texts, the voice is similar rather
than identical. OpenAI’s voice names map to the nearest preset (alloy and
nova to ava, ash and echo to leo, ballad and fable to james,
coral and shimmer to zoe, onyx and cedar to omar, sage to
nora, verse to arjun, marin to priya).
Descriptions work best when they follow this template (age, gender, accent,
pitch, timbre, pacing, emotion and intensity, and for realistic voices a
domain and role) and end with a full stop: the model reads a short or
unfinished description aloud more often. Izwi adds the final full stop when
it is missing.
Maya 2 Native
Maya2-Native-2.5B has 20 named voices. Any voice can speak any of the
model’s 12 languages: the last column lists the languages with at least an
hour of that voice’s training audio, and the voice speaks the others
cross-lingually. No built-in voice has Assamese training audio. Voice names
are not case-sensitive, and a request without one uses Ahana. OpenAI’s
voice names map to voices trained on ten or more languages (alloy and sage
to Vaidehi, ash to Siddhant, ballad and onyx to Vasant, cedar
and fable to Anirban, coral to Jahnavi, echo to Nakul, marin to
Malini, nova to Ahana, shimmer to Charvi, verse to Satvik).
Example: