AuK GGUF for audio.cpp

These GGUF files are for use with audio.cpp. The card's base_model points to the Tencent AuK base model. This package also includes optional files converted from AuK-Flash; Flash is a separate distilled variant and is not the pinned base model.

Output quality: We validate parity with upstream Python for the tested configuration, not whether AuK's output meets every quality expectation. If a result is disappointing, listen to the corresponding Python reference WAV linked below first. A similar Python result points to the upstream model's behavior, not necessarily an audio.cpp conversion issue. The quantized component combinations have not received the same 16-task parity validation.

Pass the downloaded directory to audiocpp_cli --model. Keep the config/ and tokenizer/ directories alongside the GGUF files. The default components are auk-base-f32.gguf, qwen2.5-omni-3b-bf16.gguf, and auk-vae-f32.gguf. AuK-Flash requires --session-option auk.variant=flash; its default generator is auk-flash-f32.gguf. The Qwen component is Qwen2.5-Omni-3B.

For example, instruction TTS with explicitly selected Base components:

audiocpp_cli --task tts --family auk \
  --model /path/to/AuK-Base-and-Flash-GGUF --backend cuda \
  --session-option auk.variant=base \
  --session-option auk.model_gguf=auk-base-f32.gguf \
  --session-option auk.qwen_gguf=qwen2.5-omni-3b-bf16.gguf \
  --session-option auk.vae_gguf=auk-vae-f32.gguf \
  --text "The next train leaves in ten minutes." \
  --request-option "instruct=A warm male voice speaking clear English." \
  --request-option duration_sec=4 --seed 42 --log --out speech.wav

To use AuK-Flash, select auk.variant=flash and an auk-flash-*.gguf generator. To use the Qwen Q8_0 component, select auk.qwen_gguf=qwen2.5-omni-3b-q8_0.gguf. Keep the F32 VAE in either case.

For editing, use --task gen --audio input.wav --text "<editing instruction>". See the audio.cpp AuK guide and the AuK cookbook for request options and example instructions.

Sixteen-task validation

The 16-task C++/Python comparison forced FP32 inference and disabled TF32 in both implementations. Input audio and reference voices came from the upstream AuK demo assets. Durations not shown below matched the source recording.

All 16 C++ requests completed and produced 24 kHz WAVs with the same frame counts as the Python outputs. The cosine values compare C++ WAVs with the Python baseline. Waveform cosine alone does not establish whether an edit followed its instruction.

The no-TF32 C++ run matched the Python WAVs at waveform cosine 0.999989991 or higher for all 16 tasks; the files were not byte-identical. These historical results use the original combined GGUF and are not a 16-task parity claim for the component GGUFs in this package.

Dtypes and validation scope

All component smoke tests used CUDA and the F32 VAE. A pass here means the listed task loaded the selected GGUFs and generated a WAV; it does not establish Python parity or task quality.

The Base F16 + Qwen Q8_0 and AuK-Flash Q8_0 + Qwen Q8_0 combinations were not tested.