text-to-speech

Text-to-Speech (Supertonic)

Supertonic is a fully local text-to-speech engine that converts any text into studio-quality 44.1kHz WAV audio. It runs entirely on CPU via ONNX Runtime — no GPU, no cloud API, no network needed after the initial ~404 MB model download. It synthesizes audio 3-4x faster than real-time on a typical server CPU.

System Requirements

• Python 3.8+
• ~500 MB RAM for model loading
• ~404 MB disk for model files (auto-downloaded on first use)
• No GPU required

Installation

pip install supertonic

On Ubuntu 24.04+ (PEP 668):

pip install supertonic --break-system-packages

Models auto-download from Hugging Face on first synthesis. Cached at ~/.cache/supertonic3/.

Quick Start — Python API

from supertonic import TTS
import soundfile as sf

tts = TTS()  # loads supertonic-3 (31 languages)
style = tts.get_voice_style("M1")  # male voice
audio, duration = tts.synthesize("Hello, world!", voice_style=style)
sf.write("output.wav", audio.squeeze(), tts.sample_rate)

The bundled script at scripts/synthesize.py wraps this into a one-liner:

python scripts/synthesize.py --text "Hello, world!" --output output.wav

API Reference

Initialization

tts = TTS(
    model="supertonic-3",     # "supertonic" (en), "supertonic-2" (5 langs), "supertonic-3" (31 langs)
    model_dir=None,           # override model cache directory
    auto_download=True,       # auto-download missing models
    intra_op_num_threads=None, # ONNX thread count (None = auto)
    inter_op_num_threads=None, # ONNX thread count (None = auto)
)

Synthesis

audio, duration = tts.synthesize(
    text,                     # text string (up to 100,000 chars)
    voice_style=style,        # Style object from get_voice_style()
    total_steps=8,            # quality: 5=fast, 8=balanced, 12=high
    speed=1.05,               # speech speed: 0.7 (slow) to 2.0 (fast)
    max_chunk_length=None,    # chars per chunk (default: 300, Korean: 120)
    silence_duration=0.3,     # seconds of silence between chunks
    lang=None,                # language code or None for auto
    verbose=False,            # print progress
)
# audio: numpy array shape (1, num_samples), float32
# duration: numpy array with total seconds