Expand Dataset and Improve Training Pipeline

## Overview
Expand the current dataset and improve the training pipeline to achieve better performance on the raga classification task.

## Current Dataset Status
- **Size**: 74 audio files (37 Carnatic + 37 Hindustani)
- **Accuracy**: 50% (honest CNN baseline)
- **Coverage**: Limited raga representation

## Dataset Expansion Goals
- [ ] Increase to 500+ audio files per tradition
- [ ] Ensure balanced representation across major ragas
- [ ] Add metadata for each audio file (raga, tradition, performer, etc.)
- [ ] Implement data augmentation for audio files
- [ ] Create train/validation/test splits with proper stratification

## Training Pipeline Improvements
- [ ] Implement proper cross-validation
- [ ] Add early stopping and model checkpointing
- [ ] Implement learning rate scheduling
- [ ] Add data augmentation (pitch shift, time stretch, noise)
- [ ] Implement gradient clipping and regularization
- [ ] Add comprehensive logging and monitoring

## Data Sources
- [ ] Integrate Saraga dataset (MTG professional recordings)
- [ ] Add YouTube audio processing pipeline
- [ ] Include high-quality studio recordings
- [ ] Add live performance recordings
- [ ] Ensure copyright compliance

## Success Criteria
- [ ] Dataset size >1000 audio files
- [ ] Balanced representation across traditions
- [ ] Training pipeline achieves >70% accuracy
- [ ] Proper evaluation metrics implemented
- [ ] Reproducible training process

## Files to Create/Modify
- `ml/training/dataset_expansion.py`
- `ml/training/improved_training_pipeline.py`
- `data/dataset_metadata.json`
- `docs/DATASET_EXPANSION_PLAN.md`

## Priority: High
Larger, better dataset is crucial for achieving production-quality performance.

Review the repository's current dataset and training implementation before scoping the work across ml/training/dataset_expansion.py, ml/training/improved_training_pipeline.py, data/dataset_metadata.json, and docs/DATASET_EXPANSION_PLAN.md. No test or entry point is named; done means the stated dataset, evaluation, reproducibility, accuracy, balance, metadata, and copyright goals are addressed.

Scritto dal modello di indicizzazione a partire dal testo della issue.