Generate natural, human-like speech from text using state-of-the-art TTS models.

Overview

Izwi’s text-to-speech converts written text into spoken audio. Features include:
  • Natural voices — High-quality, expressive speech
  • Local audio output — WAV for files and raw PCM for low-level API clients
  • Speed control — Adjust playback speed
  • Streaming — Real-time audio generation
  • Local processing — No cloud, complete privacy

Getting Started

Download a TTS Model

Kokoro-82M Prerequisite (espeak-ng)

Kokoro-82M is the default TTS model. Install espeak-ng on your system first. Izwi uses it for Kokoro phonemization and will return an error if it is missing.

Generate Speech

Command line:
Play after generation:
Izwi saves the file, then plays it with the system player.

Using the CLI

Basic Usage

Options

Examples

WAV output:
Adjust speed:
Read from stdin:
Streaming output:
Clone from reference audio:
Clone with Chatterbox (no transcript), more expressive:
For expressive or fast-talking reference voices, Resemble AI suggests a lower --cfg-weight (about 0.3) with --exaggeration around 0.7 or higher. Reuse a saved voice:
Design a prompted voice:

Using the Web UI

  1. Navigate to Text to Speech in the sidebar
  2. Enter your text in the input field
  3. Select a voice (if available)
  4. Click Generate
  5. Play or download the audio
With Chatterbox selected, the dialog shows sliders for exaggeration, CFG weight and min-p. With Chatterbox Multilingual it also asks which language the text is in. With Maya 2 Native it asks the same, with Auto-detect (the default) to read the language from the text’s script, and offers its inline tags. With Fish Audio Drama 3 the saved voice is optional (None (seeded voice)) and the dialog has a Seed, filled with a new number each time it opens; keep it to get the same voice again. Dialogue, beside Single voice, edits a scene:
  • Speakers: each speaks in a saved voice that has a transcript, or in a voice the seed sets. Up to 16 speakers and 8 saved voices.
  • Lines: each has its speaker, its text and Insert a tag for performance tags. Preview speaks one line without saving it.
  • Generate scene speaks every line as one take.
A Drama 3 take’s page has Edit this take: the dialog opens with the take’s text (its lines, for a dialogue), its voices and its seed. Change the line and generate: the new take continues from the old one, so the model says the new text in the context of what it already said. The new take’s page links back to it under Edited from. A dialogue take’s ⋯ menu has Open in dialogue editor, which brings its lines back as a new scene with the same voices and seed. Saved voices and the seed come back. A voice given as a reference clip does not, because a take keeps no clip audio: its speaker starts in a seeded voice. Takes made before take editing kept no voices or seed and cannot be edited; their lines still open in the dialogue editor, every speaker in a seeded voice.

Features

  • Live preview — Hear audio as it generates
  • Download — Save audio files locally as WAV, MP3, FLAC or Opus
  • History — Access recent generations

Using the API

Endpoint

Request

Response

Binary audio data with appropriate Content-Type header. Set stream to true or stream_format to sse to receive server-sent audio events instead of one binary response. See the API Reference for streaming event shapes, voice-cloning fields, saved voices, and model-specific controls.

Example (curl)


Available Models

For reference-audio cloning, use Base variants, VibeVoice-1.5B, FishAudio-S2-Pro, FishAudio-Drama-3, Chatterbox-500M, Chatterbox-Multilingual-500M or Maya2-Native-2.5B.
For built-in voice presets, use Kokoro-82M, CustomVoice variants, Voxtral-4B-TTS-2603, LFM2.5-Audio-1.5B-GGUF, LFM3-Audio-2B-GGUF or Maya2-Native-2.5B.
For prompt-based voice design, use VoiceDesign variants. Kokoro-82M requires espeak-ng to be installed separately. Voxtral-4B-TTS-2603 supports 20 preset voices and emits 24 kHz audio. Its model and bundled voice assets inherit a CC BY-NC 4.0 license. Chatterbox-500M (English, 24 kHz, MIT) speaks with its built-in default voice or clones from reference audio alone. It streams (first audio in about 2 s on an Apple M1 Pro with Metal), has no speed control or voice design, and runs slower than real time (a real-time factor of about 2.2 for one request on that Mac, where its T3 model runs in F16 and the model takes about 2.6 GB; set IZWI_METAL_DTYPE=f32 to keep it in F32), so streamed audio can pause between windows. Long text is split between sentences and spoken in full. Every utterance carries Resemble AI’s imperceptible PerTh watermark; operators can turn it off with IZWI_CHATTERBOX_WATERMARK=false. Chatterbox-Multilingual-500M does the same in Arabic, Chinese, Danish, Dutch, English, Finnish, French, German, Greek, Hebrew, Hindi, Italian, Korean, Malay, Norwegian, Polish, Portuguese, Russian, Spanish, Swahili, Swedish and Turkish, named per request with language (English when unset). Japanese is not supported yet. Maya1-3B (English, 24 kHz, Apache-2.0) speaks with a built-in voice or one described in instructions, and performs tags such as <laugh> placed in the text. On an Apple M1 Pro with Metal it streams first audio in about 2 s and runs at a real-time factor of about 4.5 for one request (about 3 with Maya1-3B-GGUF, which also needs half the memory). A request’s audio is the same alone, streamed or beside other requests for the same seed. LFM3-Audio-2B-GGUF (24 kHz, LFM Open License v1.0) speaks with a female or male voice in one of 19 locales named with language (American English when unset). It streams, splits long text into sentences automatically, and has no voice cloning, voice design or speed control. The same model also transcribes speech and answers in voice chat. FishAudio-Drama-3 (44.1 kHz, Fish Audio Community License v1.0) speaks without any voice input, or clones a reference voice; see Fish Audio Drama 3 below. For built-in speaker IDs, see Voice Presets.

Fish Audio Drama 3

FishAudio-Drama-3 needs no voice input. Without a reference voice, the sampling seed sets the voice: the same seed gives the same voice, and a request without one gets a different voice each time. With reference audio and its transcript, or a saved voice that has one, it speaks in that voice (see Voice Cloning). It reads the language from the text; a [language: xx] tag in the text steers it. Tags in the text direct the performance where they stand: [sighs], [laughs], [pause], [gasps], [clears throat], [hesitates], free-text tags such as [with quiet excitement], and <whisper>…</whisper>. They reach the model as text.
Long text is spoken as a chunk session. The model’s own chunker cuts the text into chunks of about 300 bytes, at sentence ends where it can, and each chunk is its own generation. Each chunk’s prompt holds the references and the first and latest chunks with the audio already generated for them, so the voices carry through the whole text. Chunks play one after another with no added pause. With --stream (or stream: true on the API), audio arrives as it is generated: first after 4 frames, then every 16 frames.

Dialogue

One request can voice up to 16 speakers. Mark each turn with a speaker tag, <|speaker:0|> to <|speaker:15|>, or send a script of {speaker, text} lines. Give each speaker a reference voice (a clip with its transcript, or a saved voice), up to 8 references in all; a speaker without one speaks in a voice the seed sets. The chunker keeps each turn whole with its speaker.
On the API, the same request is references and script on /v1/audio/speech or the text-to-speech records; see Drama 3 dialogue.

Take editing

A take can continue from earlier takes: to say a line again, or the next line, in the context of what was already said. Its prompt holds the earlier takes as finished turns, their text and the audio generated for it, before the new text; that is how the model was trained to edit takes. Each Drama 3 take keeps the audio codes of its first and last chunks, its voices (saved voices by id) and its seed, so a later take can name it. On the text-to-speech records, send previous_take_ids with up to 2 earlier takes, oldest first, beside the new text, voices and seed. Voices and seed are not inherited: send the ones you want, for example those the earlier take’s record reports. Takes of another model, unfinished or failed takes, and takes made before take editing are refused. See Drama 3 take editing. In the Web UI, Edit this take does this for you. Its license allows commercial use only below USD 500,000 in annual revenue, and offering the model to third parties as a hosted speech service needs a license from Fish Audio; see Models.

Audio Formats

AAC is not offered: AAC encoders need a patent licence. Studio exports and speech history downloads (?format=mp3) take the same formats.

Tips

  1. Punctuation matters — Use proper punctuation for natural pauses
  2. Break long text — Split very long text into paragraphs
  3. Test different speeds — Find the right pace for your use case
  4. Use appropriate models — Larger models = better quality but slower

See Also