Audio8-TTS-MLX-8bit

Native MLX 8-bit release of Audio8/Audio8-TTS-Preview-0.6b for Apple Silicon.

Inference code, installation, API documentation, tests, and benchmark evidence: vanch007/mlx-audio8-tts.

Artifact

• Affine 8-bit, group size 64, sensitive-bf16 policy.
• 827 MiB language-model weights; 2.08 GiB complete repository download.
• The shared 1.26 GiB neural codec, embeddings, and Fast AR depth decoder are kept at higher precision to protect speech quality.
• 44,100 Hz output, 10 acoustic codebooks.

M3 Max benchmark

Seeded post-warm-up RTF on the release checkpoint: 0.983 English, 0.922 Chinese, and 0.793 Cantonese. Model download, loading, and warm-up are excluded. Lower is better; values below 1.0 are faster than real-time. The reproducible script and report are published with the source project.

Usage

git clone https://github.com/vanch007/mlx-audio8-tts.git
cd mlx-audio8-tts
pip install -e '.[server]'

mlx-audio8-tts generate \
  --model vanch007/Audio8-TTS-MLX-8bit \
  --text "你好，欢迎使用 MLX Audio8 TTS。" \
  --output output.wav

This is an independent Apache-2.0 MLX conversion. See the upstream project for the original architecture and checkpoint.