Audio8-TTS-Preview-0.6b-MLX-4bit

MLX 4-bit (group size 32) quantization of Edge0/Audio8-TTS-Preview-0.6b, a 0.6B DualAR text-to-speech model with zero-shot voice cloning across 11 languages (Cantonese, Chinese, Dutch, English, French, German, Italian, Japanese, Korean, Polish, Spanish). Runs on Apple Silicon with mlx-audio (arktts model class).

Provenance

• Source: Edge0/Audio8-TTS-Preview-0.6b @ revision f07040f3d151f1ba0253bfb92cb2f5dd38b44594 (Apache-2.0).
• Only the language model's linear layers are quantized to 4-bit affine (gs32). The 44.1 kHz codec is kept in fp32 and the tied token-embedding table in bf16 (it doubles as the output head), so audio fidelity is bounded by the LM tier, not the vocoder.
• The upstream codec.pth (torch pickle) was re-serialised to safetensors with keys unchanged; auto_map was dropped from config.json so loading never asks for remote code.
• Built with the standalone lane pipelines/audio8_tts_mlx_release.py in ajentik/majek.

Smoke gate

Before upload this tier synthesised one sentence per language below with the default voice and was transcribed back by majentik/Qwen3-ASR-1.7B-MLX-8bit. A tier ships only if every sample is finite, > 1 s, non-silent and has character error rate <= 0.35.

Lang	Audio	RTF	CER	Verdict
en	3.7 s	0.44	0.000	pass
zh	3.9 s	0.42	0.000	pass
ko	4.7 s	0.53	0.000	pass
ja	4.7 s	0.48	0.036	pass
yue	4.3 s	0.34	0.250	pass

Peak unified memory during synthesis: 6.4 GB (mlx 0.32.2, mlx-audio 0.5.4, Apple M5 Max 128GB). Gate record: .sisyphus/evidence/audio8-tts/4bit/smoke.json in the repo above.

Usage

pip install -U mlx-audio soundfile
python -m mlx_audio.tts.generate --model majentik/Audio8-TTS-Preview-0.6b-MLX-4bit \
  --text "The quick brown fox jumps over the lazy dog." --output_path out

import mlx.core as mx, numpy as np, soundfile as sf
from mlx_audio.tts.utils import load

model = load("majentik/Audio8-TTS-Preview-0.6b-MLX-4bit")
for r in model.generate(
    text="今天天气很好，我们一起去公园散步吧。",
    # optional zero-shot cloning:
    # ref_audio="speaker.wav", ref_text="transcript of speaker.wav",
):
    sf.write("out.wav", np.asarray(r.audio.astype(mx.float32)), 44100)

Evaluation

Only the multilingual smoke gate above is run at publish time; no MOS or speaker-similarity study has been done on this tier. Upstream reports token-exact greedy parity between the MLX port and the reference implementation for the unquantized model.

Available tiers

Repo	Bits	Group size
majentik/Audio8-TTS-Preview-0.6b-MLX-8bit	8-bit	gs64
majentik/Audio8-TTS-Preview-0.6b-MLX-4bit	4-bit	gs32

No bf16 tier is published here: the source is already bf16 and mlx-community/Audio8-TTS-Preview-0.6b-bf16 covers it.

License

Apache-2.0, inherited from the source model. See the source card for the model's intended use and limitations.