mlx-community/MOSS-Audio-Tokenizer-v2-bf16

MLX conversion of OpenMOSS-Team/MOSS-Audio-Tokenizer-v2 (OpenMOSS) — the 48 kHz stereo neural audio codec behind MOSS-TTS-Local-Transformer-v1.5: causal windowed transformer encoder/decoder with a residual LFQ quantizer (32 codebooks × 1024, 12.5 frames/s, downsample_rate 3840 per channel).

Precision: bf16 (the upstream compute_dtype; the upstream file is fp32, 8.5 GB → 4.3 GB here). Single model.safetensors with the key names mlx-audio's MossAudioTokenizer loads directly (linear1/2 → ffn.0/2, in_projs.0 → in_proj, …); config.json is the upstream config minus auto_map. Round-trip verified with mlx-audio (encode → decode of a 10 s clip).

Use with mlx-audio

from mlx_audio.codec.models.moss_audio_tokenizer import MossAudioTokenizer
tok = MossAudioTokenizer.from_pretrained("mlx-community/MOSS-Audio-Tokenizer-v2-bf16")
codes = tok.encode_audio("speech.wav", num_quantizers=12)     # [frames, nq] int32
audio = tok.decode_audio_codes(codes, num_quantizers=12)       # [samples, 2] float32 @ 48 kHz

Pass audio_tokenizer_source="mlx-community/MOSS-Audio-Tokenizer-v2-bf16" to MOSS-TTS-Local-Transformer-v1.5 generation to use this mirror instead of the upstream fp32 repo.

License

Apache-2.0 (upstream weights and this conversion). Please credit OpenMOSS for the model.