Generate speech from text.

Synopsis


Description

Converts text to speech using a TTS model. The OSS server emits WAV audio for the CLI path, with voice selection and real-time streaming.

Kokoro-82M Prerequisite (espeak-ng)

Kokoro-82M requires espeak-ng to be installed on the host system (used for phonemization).

Arguments


Options

--exaggeration, --cfg-weight and --min-p are sent only when set, and the server refuses one the chosen model does not support.

Examples

Basic usage

Kokoro-82M

Play after generation

Izwi saves the generated audio, then plays it with the system player.

WAV output

Adjust speed

Read from stdin

CustomVoice speaker presets

Reference-audio voice cloning

Base, VibeVoice, FishAudio-S2-Pro and FishAudio-Drama-3 models can clone from a reference audio sample. Provide the audio and matching transcript together:
For longer transcripts, read the reference text from a file:

Chatterbox

Chatterbox-500M speaks with its built-in default voice, or clones from --reference-audio alone (a transcript is optional and ignored):
For expressive or fast-talking reference voices, Resemble AI suggests a lower --cfg-weight (about 0.3) with --exaggeration around 0.7 or higher. Chatterbox-Multilingual-500M speaks the language --language names (English without one), with the same voices and controls:
Chatterbox-500M returns 400 for any language but English. With --stream, Chatterbox’s first audio arrives after about 2 s on an Apple M1 Pro (Metal); the rest follows in windows. It refuses --speed and --instructions.

Maya1

Maya1-3B speaks with one of nine described voices (--speaker ava, the default, or leo, nora, james, …), or with a voice you describe. Tags in the text, such as <laugh>, <sigh> or <whisper>, are performed where they stand:
The tags are <angry>, <appalled>, <chuckle>, <cry>, <curious>, <disappointed>, <excited>, <exhale>, <gasp>, <giggle>, <gulp>, <laugh>, <laugh_harder>, <mischievous>, <sarcastic>, <scream>, <sigh>, <sing>, <snort> and <whisper>; the server returns 400 for any other <word> in Maya1’s text. Long text is spoken in sentences of up to about 11 seconds, each with the same voice description.

Maya 2 Native

Maya2-Native-2.5B speaks English and 11 Indic languages with one of 20 voices (--speaker Ahana, the default, or Vaidehi, Anirban, …; see Voice Presets). It reads the language from the text’s script, or takes a code with --language. Devanagari text reads as Hindi and Bengali script as Bengali, so name Marathi (mr) and Assamese (as):
It also clones a voice from reference audio and its exact transcript, and the clone can speak another of its languages. --reference-language names the language of the reference; without it, Izwi reads it from the transcript:
Tags in square brackets, such as [laughs], [sighs] or [whispers], are performed on a best-effort basis. Maya Research names [laughs], [sighs], [whispers], [sad], [excited], [curious], [angry] and [surprised]; other lowercase [tags] are passed to the model too. Numbers, dates, currency and similar are spoken as words in the text’s language. Text longer than 280 characters is spoken in whole sentences of up to 280 characters (about 20 seconds each). It refuses --speed and --instructions.

LFM3 Audio

LFM3-Audio-2B-GGUF speaks with a female or male voice (see Voice Presets) in one of 19 locales that --language names: en-US (the default), en-GB, ar, da, de, el, fi, fr, he, it, ja, ko, nl, pl, pt, es, sv, tr or zh:
Long text is split into sentences automatically, and --stream writes 24 kHz audio as it is generated. Sampling follows Liquid AI’s defaults (temperature 1.0, top-k 128); --temperature overrides the temperature. It refuses --reference-audio, --saved-voice-id, --instructions and --speed.

Fish Audio Drama 3

FishAudio-Drama-3 needs no voice input. Without a reference, --seed sets the voice: the same seed gives the same voice, and without one each run picks a new voice. Tags in the text, such as [sighs], [laughs], [pause], [gasps], [clears throat], [hesitates], free-text ones like [with quiet excitement] and <whisper>…</whisper>, direct the delivery where they stand:
It clones a voice from --reference-audio with its transcript (--reference-text or --reference-text-file), or from --saved-voice-id when the saved voice has a transcript. A reference file may last up to 60 seconds:
It reads the language from the text. Long text is spoken in chunks of about 300 bytes that carry the voice from one to the next, and --stream writes 44.1 kHz audio as it is generated. Sampling defaults to temperature 1.0 and top-p 0.9; --temperature overrides the temperature.

Dialogue

FishAudio-Drama-3 voices up to 16 speakers in one take. Write the dialogue as a script, one JSON object per line:
Give each speaker a reference voice: --speaker-reference N=AUDIO:TRANSCRIPT_FILE for a clip and a file with its transcript, or --speaker-voice N=SAVED_ID for a saved voice. Both repeat, up to 8 references, and several may voice one speaker. --reference-audio with --reference-text (or --saved-voice-id) still give speaker 0. A speaker without a reference speaks in a voice the seed sets:
Instead of --script, the text may carry the tags itself: izwi tts "<|speaker:0|> Ready? <|speaker:1|> Always." --model FishAudio-Drama-3. The CLI waits up to 900 s for each chunk the model’s chunker plans from the text (one chunk is about 300 bytes) before it gives up; the server’s own deadlines apply as well. See Drama 3 dialogue for what the server refuses.

Saved voice reuse

Saved voices created through Voice Studio or /v1/voices can be reused without resending reference audio:
Do not combine --saved-voice-id with --reference-audio or --reference-text; choose one voice source per request.

Prompt-based voice design

VoiceDesign models accept a natural-language direction prompt:

Streaming, then play

Izwi saves the streamed audio, then plays it.

Audio Formats


Models


See Also