README.md

# Audio Source Separation Blind Spots for Musical Instrument Segmentation

This dataset documents blind spots discovered in an audio source separation model when tested on various instrument isolation tasks.

## Model Tested

**SAM-Audio Base** (Segment Anything Model for Audio)
Model: [facebook/sam-audio-base](https://huggingface.co/facebook/sam-audio-base)

SAM-Audio is a foundation model for isolating sounds in audio using text, visual, or temporal prompts.
I have evaluated the model’s **text prompting capability for instrument isolation**.

---

## Blind Spot Results

The results are provided in [dataset.csv](dataset.csv), with corresponding audio files located in [assets](assets).


## Fine-Tuning Strategy

Instead of fixing exact dataset sizes from the start, the approach is iterative:

1. **Create datasets within a reasonable range of examples** for each problem category.
2. **Fine-tune the model on this initial dataset.**
3. **Evaluate results and identify remaining failure cases.**
4. **Expand the dataset and repeat the process.**

This iterative process helps focus data collection on the most impactful failure modes.

### Dataset Categories

1. Percussion Disambiguation (~3k–6k samples)
Examples where similar percussion instruments occur together, such as snare, crash cymbals, and hi-hats.
The goal is to improve the model's ability to distinguish between closely related percussive sounds.

2. South Asian Instruments (~800–2k samples)
Examples featuring instruments such as sitar, sarangi, madal, and tabla across different musical contexts.
These instruments are underrepresented in many training datasets and may require targeted data collection.

3. Pitch-Similar Instruments (~2k–5k samples)
Cases where instruments with overlapping pitch ranges appear together (e.g., vocals with guitar solos, strings with woodwinds).
These cases test the model’s ability to separate sources that share similar spectral characteristics.

4. Complex Mixes (~3k–5k samples)
Audio clips containing multiple overlapping instruments or dense arrangements that create difficult separation scenarios.


To prepare datset, larger variant of sam_audio can be used to assist with dataset labeling and segmentation, and then fine-tune a base model for the specific task of musical instrument separation.