audio-cpp/AuK-Base-and-Flash-GGUF

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.

Task	Tested instruction / setting	C++ vs Python WAV cosine	Listen
Zero-shot TTS	Reference voice; target sentence; 6 s	0.999999992	Python / C++
Instruct TTS	Calm woman speaking clear English; 3 s	0.999999898	Python / C++
Speech content editing	Replace "but accepting what we cannot have" with "and living well with dreams unmet"; 7 s	0.999999982	Python / C++
Lyric editing	Replace "rear view" with "like you" in isolated vocals	1.000000000	Python / C++
Pitch editing	Raise pitch by 2 semitones	0.999989991	Python / C++
Speed editing	Speed 1.5x; 6.86 s	0.999999993	Python / C++
Volume editing	Increase volume by 10 dB	0.999999997	Python / C++
Emotion editing	Change emotion to happy	0.999999993	Python / C++
Timbre editing	Change to a deep, calm male voice	0.999999991	Python / C++
De-accent	Change Sichuan-accented speech to standard Mandarin	0.999999795	Python / C++
Nonverbal editing	Add a cough before "We tested"; 10.44 s	0.999991188	Python / C++
Whisper conversion	Speak the source in a quiet whisper	0.999998371	Python / C++
Speech enhancement	Preserve speakers; remove noise and reverberation	0.999999994	Python / C++
Speech separation	Keep the second speaker to start talking	0.999999985	Python / C++
Music separation	Keep the singing voice; remove other audio	0.999999998	Python / C++
Target speaker extraction	Keep the speaker saying "get what"	0.999999997	Python / C++

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.

Generator	Qwen conditioner	Tested task(s)
Base F32	BF16	Zero-shot TTS; instruct TTS
Base F16 or Q8_0	BF16	Instruct TTS, one run per dtype
Base F32 or Q8_0	Q8_0	Zero-shot TTS, one run per combination
AuK-Flash F32	BF16	Zero-shot TTS; instruct TTS
AuK-Flash F16 or Q8_0	BF16	Instruct TTS, one run per dtype
AuK-Flash F32	Q8_0	Zero-shot TTS, one run
AuK-Flash F16	Q8_0	Instruct TTS, one run

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