Crate kittentts 

kittentts

Rust port of KittenTTS — an ultra-lightweight ONNX-based text-to-speech engine.

Quick start

use kittentts::{KittenTTS, download};

// Download the model from HuggingFace (cached after first run)
let tts = download::load_from_hub("KittenML/kitten-tts-mini-0.8").unwrap();

// Generate audio samples (Vec<f32>, 24 kHz mono)
let audio = tts.generate("Hello from Rust!", "Jasper", 1.0, true).unwrap();

// Or write directly to a WAV file
tts.generate_to_file(
    "Hello from Rust!",
    std::path::Path::new("output.wav"),
    "Jasper",
    1.0,
    true,
).unwrap();

Mobile (iOS / Android)

Phonemisation uses the pure-Rust espeak-ng crate with bundled data. The same generate API works on every platform with no extra setup — no C library, no system dependencies.

You can also skip phonemisation entirely and pass pre-computed IPA:

use kittentts::{KittenTTS, download};

let tts = download::load_from_hub("KittenML/kitten-tts-mini-0.8").unwrap();
let audio = tts.generate_from_ipa("həloʊ fɹʌm ɹʌst", "Jasper", 1.0, 20).unwrap();

Build requirements

Platform	Requirement
All platforms	None — the espeak-ng crate is pure Rust with bundled data

Pipeline (matches Python implementation)

1. Text preprocessing — numbers, currencies, abbreviations → spoken words.
2. Chunking — long texts split into ≤ 400-char sentence chunks.
3. Phonemisation — pure-Rust espeak-ng converts text to IPA phonemes.
4. Tokenisation — IPA characters mapped to integer token IDs.
5. ONNX inference — model takes (input_ids, style, speed), outputs audio.
6. Tail trim — last 5 000 samples removed (silence artifact).
7. Concat — per-chunk audio concatenated into a single waveform.

Re-exports

pub use model::KittenTtsOnnx as KittenTTS;

pub use model::SAMPLE_RATE;

pub use encoding::AudioEncoder;

pub use encoding::AudioFormat;

pub use encoding::EncoderFactory;

Modules

download

HuggingFace Hub model downloader — mirrors get_model.py.

encoding

Audio encoding — convert raw f32 samples to various audio formats.

ffi

C FFI — bridges KittenTtsOnnx to iOS / Android callers.

model

ONNX model runner — mirrors Python’s KittenTTS_1_Onnx.

npz

Minimal NPZ / NPY loader.

phonemize

Phonemisation using the pure-Rust espeak-ng crate.

preprocess

Text preprocessing pipeline — mirrors kittentts/preprocess.py.

tokenize

Character-level tokeniser — mirrors Python’s TextCleaner.