Audio8-TTS-Realistic-Air-MLX-8bit-selective

Native Apple Silicon MLX checkpoint converted from Edge0/Audio8-TTS-Realistic-Air at revision 23c305cc013fde378edb459f538317043c993a74. Tested on Apple M5 24 GiB, 3 November 2026. This is a derivative conversion, not an official upstream release.

8-bit affine, group size 64, round-to-nearest, selective weights. 140 slow/fast transformer block Linear modules at 8 bits. Codec remains FP32; embeddings, output heads, norms and auxiliary projections remain BF16.

Complete checkpoint: 2,160,646,598 bytes (2.161 decimal GB), including protected dense components; 15.3% smaller than the complete source checkpoint. This is not uniform 8-bit quantization of every component.

Measured result

Five English/Spanish/bilingual utterances: 0.521 RTF versus 0.679 unquantized. Includes codec decoding; excludes model load and warm-up.

See BENCHMARKS.md for the method, per-sample/per-document timings, limitations, and companion release. benchmark-results.json contains exact measurements; quantization-audit.json records every tensor's shape/type and quantized module. Original email documents and OCR text are excluded.

Native MLX usage

Validated versions: MLX 0.32.4, MLX-VLM 0.7.11, MLX-Audio 0.6.4, Transformers 5.22.1. Use the native MLX loader rather than Transformers AutoModel. The supplied benchmark requirements include Torch for original codec conversion; Torch is not used for model inference in the converted Audio8 path.

from mlx_audio.tts.utils import load_model
import soundfile as sf
model = load_model('groxaxo/Audio8-TTS-Realistic-Air-MLX8-Selective')
for result in model.generate(text='Hola. This is a bilingual test.', seed=20261103, max_tokens=1024):
    sf.write('speech.wav', result.audio, result.sample_rate)

Reproduction and provenance

benchmark_mlx.py records the exact quantization predicates and test inputs. Audio8 samples are synthetic, generated without reference-audio cloning. No MOS or human listening score was measured. The OCR sample contains only one/two-page documents and does not test hundred-page/unlimited-horizon behavior. Allocator memory is not full system RSS. Source license and attribution remain applicable; preserve them when redistributing.