ptts for Python

Python bindings for the Phonon Rust runtime.

To try it with uv, run:

uvx ptts --lang en "Hello world" -o out.wav

uvx handles the isolated Python environment and package dependencies. The first run downloads the default model, Phonon 31m, and caches it. --model accepts a local checkpoint directory or a Hugging Face repo as hf://owner/name, such as hf://gradium/phonon-40m-fr with --lang fr.

To use the Python API in your own environment, install the package:

pip install ptts

The wheel contains no model weights. Without a model argument, TTS downloads the default model on first use and caches it:

import ptts

tts = ptts.TTS(lang="en")
print(tts.voices)
tts.save("out.wav", "Hello world")

For a source install, use uv add /path/to/xn-ptts/ptts-pyo3 or pip install /path/to/xn-ptts/ptts-pyo3. Source builds need Rust installed; supported CPython wheels do not. See the development guide for repository build commands.

model, the first argument, names a local directory, a config.json, or a Hugging Face repo as "hf://owner/name" or "hf://owner/name@revision". Left out, it is ptts.DEFAULT_MODEL. Local paths download nothing. Gated repos, such as the Phonon 40m models, need their terms accepted on the model's Hugging Face page, then HF_TOKEN set or hf auth login. Rust examples, Python, and both servers use the shared checkpoint resolver. Every checkpoint supplies its own config.json and tokenizer.json. q8 prefers model.q8.gguf and q4k model.q4k.gguf; other formats prefer model.safetensors. Voices come from voices/ or embeddings/, plus default-voice.safetensors as default. Hub downloads use the revision after @ for all files. config= and revision= still work but are deprecated.

lang is required: en, fr, de, es or pt picks how numbers, symbols and abbreviations are spelled out, ar, hi, ko, pl, sv, fi, no, it, nl, ja, zh, da and tr spell out symbols only; none uses the text as written. ptts.normalize("It costs $5.", lang="en") shows that reading, "It costs 5 dollars.", without loading a model, and takes the same rewrites.

Speech and voices

Reuse tts across requests. Use synth for an array, save for a WAV, or stream for chunks as they are decoded:

voice = tts.voices[0]
pcm = tts.synth("Hello", voice=voice)  # 1-D float32 NumPy array
seconds = tts.save("out.wav", "Hello", voice=voice)
with tts.stream("A longer sentence.", voice=voice) as audio:
    for chunk in audio:
        print(chunk.shape)  # process each PCM chunk as it arrives

tts.sample_rate is the PCM sample rate; save writes a mono 16-bit WAV and returns its duration. Leaving the with block stops a stream early and waits for its workers to finish. Streaming keeps bounded audio and latent buffers: leaving a stream unread pauses generation once they fill, and consuming chunks lets it resume.

tts.voices lists the voices that were found, and tts.languages the languages the checkpoint's config says it was trained for. When no voice is given, the checkpoint's configured default is used, then default if present, then the first by name. Pass voice="name" to any speech method to select one. A name the checkpoint does not have is looked up in the voice catalog, by name in any case and with or without accents, or by its voice ID. Only that voice is downloaded, on first use, and then joins tts.voices. The catalog is ptts.DEFAULT_VOICE_CATALOG, gradium/phonon-voices, unless voice_catalog= names another: an hf://owner/name[@revision] repo, or a local voices.jsonl or a folder holding one. tts.add_voice("NAME", "voice.safetensors") adds a voice file, such as one from the Gradium Studio scripts. ptts.catalog_voices() lists it without loading a model, and ptts.catalog_voices(catalog) another one. To change the speaking rate, pass conditions={"duration_delta": -0.3} to speak faster, or a positive value to speak slower.

Command line

The package also provides the ptts command:

python -m ptts --lang en "Hello world" -o out.wav

--model loads another model, as TTS's first argument does. --voice selects a voice of the checkpoint or of --voice-catalog, and --list-voices prints the catalog's voices. Run with --help for the remaining options. The package ships type stubs and py.typed.

Wheels

Linux and Windows x86_64 wheels target x86-64-v3 (AVX2, FMA and F16C); older x86 CPUs need a source build. The macOS x86_64 wheel does not need AVX.

Released wheels are cp39-abi3, so one per platform covers every CPython from 3.9 on, and a new CPython release needs no new wheel. Free-threaded CPython and PyPy cannot load an abi3 extension, so those two install from the sdist and compile the Rust runtime, which needs a Rust toolchain and takes a few minutes.

License

MIT or Apache-2.0, at your option. The model weights are not part of the package and come with their own licenses: Phonon 31m is under Apache 2.0, and Phonon 40m and Phonon 40m multilingual are under CC BY-NC 4.0. See each model's page on Hugging Face.