linnet-tts (Python)

Python package for the LinnetTTS native runtime: lightweight on-device text-to-speech at 24 kHz, streaming, fp16 weights, no PyTorch. The extension module is built from the Rust crate runtime/linnet-tts with maturin (abi3, one wheel per platform for every CPython >= 3.9).

Install

pip install linnet-tts soundfile      # soundfile only for writing WAV files in the snippets

Wheels on PyPI: manylinux 2.17 x86_64 and aarch64, macOS 11+ on Apple silicon; no Windows, Intel-macOS or source distribution in 0.1.0 (build from the repository on other platforms, see "Build"). The model files are not in the wheel: download the weights release https://github.com/VoxRT/LinnetTTS/releases/tag/weights-v1 (the fp16 weights, tokenizer.model and a voice .npy). They are licensed under the Linnet Weights License (https://github.com/VoxRT/LinnetTTS/blob/main/WEIGHTS_LICENSE.md), not Apache-2.0. Source, docs and the streaming contract: https://github.com/VoxRT/LinnetTTS.

import linnet_tts, soundfile as sf

model = linnet_tts.Model("linnet-en-35m.fp16.safetensors", "tokenizer.model")
voice = model.load_voice("voice.npy")                     # [T, 48] float32 .npy

audio = model.synthesize("Hello there.", voice, temp=0.3, cfg=2.0, seed=0)
sf.write("hello.wav", audio, model.sample_rate)           # float32, 24 kHz

session = model.session(voice, seed=0)
session.prime("Streaming gives the first 50 ms of audio after one step.")
for chunk in session:                                     # 80 ms per LM step
    play(chunk)

Voice.from_array(np.ndarray) wraps in-memory [T, 48] latents. The chunk sizes follow the streaming contract (docs/streaming-contract.md in the repository): the first chunk is samples_per_step - output_delay samples, later ones samples_per_step (1920), the last one output_delay (720) zeros, and the concatenation of all chunks is sample-for-sample identical to synthesize with the same parameters and seed.

Build

cd bindings/python
maturin build --release                     # wheel in target/wheels/
maturin develop --release                   # or: install into the active venv

The release profile uses fat LTO and strips symbols; the wheel is ~0.4 MB. Runtime dependency: numpy.

Test

With the wheel from PyPI: clone the repository, install pytest, and run pytest bindings/python/tests from its root (the tests import the installed linnet_tts). The tests need the model assets (gitignored). Point LINNET_ASSETS at a directory holding linnet-en-35m.fp16.safetensors, tokenizer.model and voices/linnet-en-f1.npy (default: ../../local; local/ = the unpacked weights release, see the root README at https://github.com/VoxRT/LinnetTTS):

uv venv /tmp/linnet-venv && . /tmp/linnet-venv/bin/activate
uv pip install target/wheels/linnet_tts-*.whl pytest
pytest tests

A plain python -m venv /tmp/linnet-venv followed by pip install target/wheels/linnet_tts-*.whl pytest works just as well where uv is not installed.