vanch007/mlx-pocket-tts

MLX Pocket TTS 8-bit

Ready-to-run Apple Silicon MLX weights for Pocket TTS. The generation FlowLM is dynamically quantized to 8-bit with group size 64; Mimi remains at source precision. The artifact includes 26 preset voice embeddings and supports arbitrary reference-audio voice cloning through the companion MLX implementation.

• Source code and complete usage guide: vanch007/mlx-pocket-tts
• This model: vanch007/mlx-pocket-tts
• Upstream code: kyutai-labs/pocket-tts
• Upstream weights: kyutai/pocket-tts
• Voice provenance: kyutai/tts-voices

Install and generate

Python 3.10+ and an Apple Silicon Mac are required.

git clone https://github.com/vanch007/mlx-pocket-tts.git
cd mlx-pocket-tts
uv sync
uv run mlx-pocket-tts generate \
  --model vanch007/mlx-pocket-tts \
  --voice alba \
  --text "Hello from Pocket TTS on Apple Silicon." \
  --output outputs/hello.wav

Use --ref-audio reference.wav for voice cloning, --stream for incremental generation, or mlx-pocket-tts serve --model vanch007/mlx-pocket-tts for the Web UI and streaming HTTP API. The GitHub README documents the Python API, reusable voice export, conversion, training, checkpointing and distillation surfaces.

Artifact contents

File	Purpose
model.safetensors	Strict-load MLX model weights; 8-bit FlowLM and source-precision Mimi
config.json	Architecture and quantization configuration
tokenizer.model	SentencePiece tokenizer
embeddings/*.safetensors	26 bundled preset voice states
quantization.json	Reproducible quantization manifest

Local M3 Max verification

The release artifact was tested directly after quantization:

Check	Result
Strict weight audit	pass: 316 parameter groups; no missing, extra or mismatched tensors
Generated audio	pass: finite 24 kHz waveform, 3.52 seconds
Real-time factor	0.338
Peak memory	0.355 GB
Whisper replay	WER 0.0

The source repository contains 33 automated tests and the full feature/evidence matrix. Metrics are single-machine release checks, not universal quality claims.

Provenance, license and responsible use

This conversion is derived from kyutai/pocket-tts model revision 492522650173a0653b7575cdc25ae09810e5d741. Model weights, tokenizer and bundled voice assets are distributed under their upstream terms; the Hugging Face metadata declares CC BY 4.0. Attribute Kyutai and the upstream Pocket TTS project when redistributing or publishing outputs where the license requires it. The MLX implementation itself is MIT licensed in the GitHub repository.

Use voice cloning only with the speaker's explicit authorization. Do not use this model for impersonation, deception, fraud, harassment, privacy violations, unlawful activity or other harmful purposes. Users are responsible for complying with applicable laws and the upstream terms.