fine-tuned-musicgen-small

# 🎵 Fine-Tuned MusicGen-Small (CLAP-Aligned)
### Smarter, more accurate text-to-music generation — fast, lightweight, and production-ready.

This model is a **CLAP-guided fine-tuned version of `facebook/musicgen-small`**, designed to generate music that more closely matches your text prompts.  
It is **compact**, **fast**, and ideal for **web apps, mobile apps, and real-time audio generation systems**.

---

## Why This Model?
Compared to the original MusicGen-Small, this version delivers:

- **Better alignment** between text prompts and generated music  
- **More stable pitch** and cleaner melodic structure  
- **Faster inference** and smaller size than MusicGen-Large  
- **Easy integration** into any backend (FastAPI, Node.js, Python servers)

---

## Quick Metrics (30-example evaluation)

| Model                     | FAD_CLAP ↓ | CLAP Score ↑ | Pitch Stability ↑ |
|---------------------------|-----------:|--------------:|-------------------:|
| Base MusicGen-Small       | 0.9726     | 0.2885        | Low                |
| **Fine-tuned (this model)** | **0.9093** | 0.3076        | **High**           |
| MusicGen-Large (SOTA)     | 0.9269     | **0.3266**    | Medium             |

**Highlights:**
- ⚡ **Lowest FAD_CLAP** — closest to real audio distribution  
- 🎧 Strong boost in **semantic match** between caption & audio  
- 🎼 **Most stable pitch** among all compared models  

---

## Generate Music in 5 Lines

```python
from transformers import AutoProcessor, MusicgenForConditionalGeneration
import soundfile as sf

model_id = "Bashaarat1/fine-tuned-musicgen-small"
processor = AutoProcessor.from_pretrained(model_id)
model = MusicgenForConditionalGeneration.from_pretrained(model_id)

inputs = processor(text=["Lo-fi beat with soft piano"], return_tensors="pt")
audio = model.generate(**inputs)[0, 0].cpu().numpy()
sf.write("music.wav", audio, 16000)

---

Drop-in Backend Integration

pip install fastapi uvicorn transformers torch soundfile

Example:

@app.post("/generate")
def generate_music(prompt: str):
    inputs = processor(text=[prompt], return_tensors="pt")
    audio = model.generate(**inputs)[0, 0].cpu().numpy()
    ...

This makes deployment into web apps, mobile apps, Unity games, or cloud APIs extremely simple.

---

Ideal Use Cases

• AI music assistants
• Creative tools & production apps
• Background music generation (games, videos, VR)
• Generative music prototypes and research

---

Notes

• Designed for short music clips (5–10 seconds).
• Works best with English prompts.
• Not intended for imitating specific artists.

---

Acknowledgments

This work builds on Meta’s MusicGen and LAION’s CLAP models.

---

Citation

@misc{fine_tuned_musicgen_small_2025,
  title        = {Fine-tuned MusicGen-Small (CLAP-Aligned)},
  author       = {Bashaarat et al.},
  year         = {2025},
  howpublished = {\url{https://huggingface.co/Bashaarat1/fine-tuned-musicgen-small}}
}