audiohacking/dasheng-audiogen-gguf

Dasheng-AudioGen GGUF

GGUF-converted weights for mispeech/Dasheng-AudioGen

Dasheng-AudioGen is a unified audio generation model that can jointly synthesize intelligible speech, music, sound effects, and environmental acoustics from text descriptions.

Model Variants

Variant	DiT Model	Total Size	Quality
F16 (default)	dit.gguf	~5.7GB	Best
Q8	dit_Q8_0.gguf	~3.8GB	Great
Q4	dit_Q4_0.gguf	~2.8GB	Good

All variants include the same T5 encoder, vocoder, and tokenizer.

Files

File	Description	Size
t5_encoder.gguf	T5 text encoder (F32)	1.3GB
dit.gguf	DiT backbone (F16)	4.1GB
dit_Q8_0.gguf	DiT backbone (Q8 quantized)	2.2GB
dit_Q4_0.gguf	DiT backbone (Q4 quantized)	1.2GB
dit-multi.gguf	DiT Multilang (F16)	4.1GB
dit-multi_Q8_0.gguf	DiT Multilang (Q8 quantized)	2.2GB
dit-multi_Q4_0.gguf	DiT Multilang (Q4 quantized)	1.2GB
vocoder.gguf	Vocoder decoder (F16)	332MB
spiece.model	SentencePiece tokenizer	792KB

Usage with audiogen.cpp

# Clone and build
git clone https://github.com/audiohacking/audiogen.cpp
cd audiogen.cpp
git submodule update --init --recursive
make metal  # or: make cpu

# Download models (choose one)
make download-models      # F16 (~5.7GB) - best quality
make download-models-q8   # Q8 (~3.8GB) - great quality
make download-models-q4   # Q4 (~2.8GB) - good quality

# Generate audio
./build-metal/dasheng-audiogen \
    models/t5_encoder.gguf models/dit.gguf models/vocoder.gguf models/spiece.model \
    --caption "A dog barking loudly" --output output.wav

Prompt Format

Supports the same prompt tags as the original model:

Flag	Tag	Description
--caption	<|caption|>	Main audio description (required)
--speech	<|speech|>	Speech characteristics
--asr	<|asr|>	Text to speak
--sfx	<|sfx|>	Sound effects
--music	<|music|>	Music description
--env	<|env|>	Environment/ambience

Example: Complex multi-tag prompt

./build-metal/dasheng-audiogen \
    models/t5_encoder.gguf models/dit.gguf models/vocoder.gguf models/spiece.model \
    --caption "A gritty detective narrating" \
    --speech "gritty deep male voice" \
    --asr "The city never sleeps, but it sure knows how to cry." \
    --sfx "heavy rain hitting pavement" \
    --music "melancholic solo saxophone" \
    --env "distant urban ambience" \
    --output noir_detective.wav

Generation Options

Option	Default	Description
--steps N	25	Number of diffusion steps
--duration SECS	10	Audio duration in seconds
--cfg SCALE	3.0	Classifier-free guidance scale
--sway COEF	-1.0	Sway sampling coefficient
--seed N	random	Random seed for reproducibility
--threads N	4	CPU threads

Performance

Backend	25 Steps	Hardware
Metal GPU	~6s	Apple M3 Ultra
CPU	~163s	Apple M3 Ultra (16 threads)

Original Model

This is a GGUF conversion of mispeech/Dasheng-AudioGen. Please refer to the original model card for more details about the model architecture and training.

License

Apache-2.0 - Same as the original model