README.md

---
title: Dasheng Mix Audiogen
emoji: 🏃
colorFrom: green
colorTo: indigo
sdk: gradio
sdk_version: 6.5.1
app_file: app.py
pinned: false
license: apache-2.0
short_description: Generate mix audio contains audio,speech and music.
python_version: "3.10.13"
models:
  - mispeech/Dasheng-AudioGen
preload_from_hub:
  - mispeech/Dasheng-AudioGen
  - mispeech/dashengtokenizer
  - google/flan-t5-large config.json,model.safetensors,spiece.model,tokenizer.json,tokenizer_config.json,special_tokens_map.json
---

# Dasheng AudioGen Space

This Gradio Space loads
[`mispeech/Dasheng-AudioGen`](https://huggingface.co/mispeech/Dasheng-AudioGen)
and performs audio generation on the Space GPU. It does not call an external
audio-generation API.

## Hugging Face deployment

1. Select **ZeroGPU** or a dedicated CUDA GPU in the Space hardware settings.
   ZeroGPU requires a Hugging Face PRO account for personal Spaces.
2. Push this repository to the Space. The model, audio tokenizer, and Flan-T5
   text encoder are preloaded during the build.
3. Configure Prompt Refiner secrets if Auto Mode is required:
   - Default XI mode: `XI_API_KEY`; optionally `XI_MODEL_NAME`.
   - OpenAI-compatible mode: `PROMPT_REFINER_MODE=openai`, `API_KEY`,
     `MODEL_NAME`, and `LLM_BASE_URL`.
   - CLAW mode: `PROMPT_REFINER_MODE=claw` and `CLAW_API_URL`.

Manual Mode does not require a Prompt Refiner secret.

## Local test

The model has been tested in the `dsag` Conda environment:

```bash
conda activate dsag
pip install "gradio==6.5.1" "openai>=1,<3"
python app.py
```

On the current AArch64 test host, preload scikit-learn's bundled OpenMP library
if Python reports `cannot allocate memory in static TLS block`:

```bash
export LD_PRELOAD="$CONDA_PREFIX/lib/python3.11/site-packages/scikit_learn.libs/libgomp-a49a47f9.so.1.0.0"
python app.py
```

The local model runtime can be adjusted with these optional variables:

| Variable | Default | Description |
|---|---:|---|
| `AUDIOGEN_MODEL_ID` | `mispeech/Dasheng-AudioGen` | Hub model ID or local snapshot path |
| `AUDIOGEN_DEVICE` | `cuda` | PyTorch inference device |
| `AUDIOGEN_DTYPE` | `float32` | `float32`, `float16`, or `bfloat16` |
| `AUDIOGEN_NUM_STEPS` | `25` | Denoising steps |
| `AUDIOGEN_GUIDANCE_SCALE` | `5.0` | Classifier-free guidance scale |
| `AUDIOGEN_SWAY_SAMPLING_COEF` | `-1.0` | Sway sampling coefficient |
| `AUDIOGEN_LOCAL_FILES_ONLY` | `0` | Set to `1` for offline cache-only loading |
| `AUDIOGEN_OUTPUT_DIR` | `outputs` | Generated WAV directory |