pocket-tts-mlx 0.2.1

MLX backend for pocket-tts with Apple Silicon optimization

pocket-tts-mlx

MLX backend for pocket-tts optimized for Apple Silicon.

Runtime is torch-free. Torch is only required for optional parity tests.

Installation

PyPI install:

pip install pocket-tts-mlx

Local development:

pip install -e .

Model weights are downloaded from Hugging Face on first run. For voice cloning weights, accept the model terms and authenticate:

hf auth login

Quickstart

from pocket_tts_mlx import TTSModel

model = TTSModel.load_model()
state = model.get_state_for_audio_prompt("marius")
audio = model.generate_audio(
    state,
    "Hello from MLX!",
    max_tokens=200,
    warmup_frames=1,
    trim_start_ms=40,
    fade_in_ms=15,
)

CLI

Basic usage:

pocket-tts-mlx "Hello, world!" --voice marius --output output.wav

Cleaner onset (recommended if startup artifacts are audible):

pocket-tts-mlx "Hello, world!" --voice marius --output output.wav --warmup-frames 1 --trim-start-ms 40 --fade-in-ms 15

Onset Cleanup Options

• --warmup-frames: decode and discard initial Mimi frames to reduce decoder startup transients.
• --trim-start-ms: trim milliseconds from start of output.
• --fade-in-ms: apply linear fade-in at start.

Equivalent Python args are warmup_frames, trim_start_ms, and fade_in_ms.

Performance Note

generate_audio() now materializes generated chunks before returning, so np.array(audio) overhead should be near zero in normal usage.

Voices

Predefined voices:

• alba
• marius
• javert
• jean
• fantine
• cosette
• eponine
• azelma

Requirements

• Python 3.10+
• Apple Silicon Mac (M1/M2/M3/M4)
• MLX
• Internet access for initial model downloads

Notes

• Voice cloning requires Hugging Face access to kyutai/pocket-tts.
• Non-voice-cloning weights are used automatically when voice cloning is unavailable.