ilintar/moss-voicegen-gguf

MOSS-VoiceGenerator BF16 GGUF for OpenMOSS

GGUF weights for OpenMOSS-Team/MOSS-VoiceGenerator, converted for the OpenMOSS C++/GGML runtime.

Precision correction

Both files in this revision are FP16-free. The VoiceGenerator checkpoint is trained and loaded as BF16 by the official CUDA reference. FP16 is unsafe for this model because its attention-sink activations can exceed FP16's finite range.

The audio sidecar has also been rebuilt without the converter's former BF16/FP32 to FP16 intermediate cast. Ordinary learned weights are BF16. Normalization tensors and the audio tokenizer's float32 residual-vector-quantizer tensors remain FP32.

Required files

File	Tensor inventory	SHA-256
moss-voicegen.gguf	198 BF16, 113 FP32, 0 FP16	b5637c2e362940ce7680e298c36c9b600193f50396e5bb2f3a0fc73c00fcbe92
moss-voicegen.extras.gguf	586 BF16, 1,046 FP32, 0 FP16	e7b3c2964d105e1b8348219f4c840ec2710865c33b2d251edefeb9920f300873

Both files must remain beside each other. The sidecar identifies the model as moss_tts_delay, with 16 RVQ codebooks and 24 kHz mono output.

Source revisions

• VoiceGenerator: OpenMOSS-Team/MOSS-VoiceGenerator@97521ec2b6f3ec5026ac1f5751f8fc302d82c2d4
• Audio tokenizer: OpenMOSS-Team/MOSS-Audio-Tokenizer@3cd226ba2947efa357ef453bcad111b6eafba782
• Converter/runtime base: pwilkin/openmoss@d1f2976a89fd51bf949d28320af401e523b78c67, plus BF16-preserving conversion and BF16/FP32 codec-sidecar loading

Runtime requirement

Use an OpenMOSS build with BF16/FP32 codec-sidecar loading support. Older builds that require moss.codec.quantizer.* tensors to be FP16 cannot decode this corrected sidecar and fail with an expected f16 error. Converting those tensors back to FP16 is not a valid workaround because the upstream audio tokenizer keeps its residual-vector-quantizer math in FP32.

Validation

• Tensor audit: zero FP16 tensors in both files.
• CPU: model load, natural end-of-speech, 62-frame codec decode, and a 4.96-second WAV.
• Vulkan: full model/sidecar load, natural end-of-speech, 52-frame codec decode, and a 4.16-second WAV.
• CPU Whisper check of the pilot recovered the requested pangram through its final word; waveform QA found no clipping or late crop transient.

OpenMOSS usage

moss-tts-cli \
  --model moss-voicegen.gguf \
  --template voicegen \
  --text "The quick brown fox jumps over the lazy dog." \
  --instruction "Clear, neutral voice with an even tempo and precise articulation." \
  --output voice.wav

Use the dedicated VoiceGenerator template. For upstream-equivalent behavior, do not add a language field, forced duration, or hard audio-frame cap unless the application requires one.

License

The source model and these converted weights are provided under Apache 2.0.