Raga Recognition using CNN

A raga is far more than a scale: it is a melodic framework defined by characteristic phrases (pakad), ornamentation (meend, gamak) and note hierarchies that even trained listeners take years to internalize. Automating raga identification is a classic music-information-retrieval problem and an excellent deep-learning project, because the signal is genuinely hard — near-neighbour ragas share the same notes and differ only in how they move between them. The open Saraga dataset from the CompMusic project provides annotated Hindustani recordings with raga labels. This project trains a compact 2D CNN on log-mel spectrograms and chroma features to classify 30-second clips into 10 ragas, served through a demo app with confidence-ranked predictions and an honest, fully-logged evaluation.

1. Saraga Hindustani recordings are segmented into 30-second clips, split at the recording level into train, validation and test sets.
2. Each clip is converted to a 128-band log-mel spectrogram plus tonic-normalized 12-bin chroma features.
3. Training clips pass through augmentation (time/frequency masking, ±1 semitone pitch shift) into the 2D CNN with cross-entropy loss.
4. The network trains for about 40 epochs with Adam, cosine decay and early stopping; every epoch logs accuracy, macro-F1 and loss curves.
5. At inference the demo app computes the same features for a clip, runs the trained weights and displays top-5 raga probabilities with confidence bars.
6. The notebook's confusion matrix shows which near-neighbour raga pairs confuse the model, forming the report's error analysis.

Tech stack:

• Python 3.10, PyTorch (CNN training and inference)
• librosa (audio loading, mel spectrograms, chroma)
• scikit-learn (metrics: accuracy, F1, confusion matrix)
• Matplotlib, Seaborn (training curves, per-class analysis)
• Saraga Hindustani collection subset (open dataset, CompMusic)
• Flask / static demo app (prediction interface)

• [Saraga dataset pipeline] Recording-level train/validation/test splits on a 10-raga subset of the open Saraga Hindustani collection — no clip leakage between splits
• [Tonic-normalized features] 128-band log-mel spectrograms plus 12-bin chroma, normalized to each recording's annotated tonic (Sa)
• [Compact 2D CNN] Three convolutional blocks plus dilated conv layers, time-axis pooling and a 10-way softmax — about 1.2M parameters
• [Audio augmentation] SpecAugment-style time/frequency masking and ±1 semitone pitch shifting during training
• [Full evaluation logging] Accuracy, macro-F1, per-class precision/recall and confusion matrix computed on held-out recordings
• [Demo prediction app] Clip selection, chromagram heatmap visualization and top-5 raga predictions with confidence bars
• [Error analysis] Confusion matrix documenting exactly which raga pairs the model confuses — the honest core of the report
• [Configurable class set] The 10-raga subset is defined in one place; adding ragas with enough recordings is a documented procedure

• Complete source code (preprocessing, training, evaluation, demo app)
• Jupyter training and evaluation notebook (buyer-run: prepare data, train, evaluate)
• Trained CNN weights from the reference training run
• Project report PDF (background, dataset analysis, model design, evaluation, error analysis)
• PPT presentation for final review
• Viva Q&A preparation document (CNNs, spectrograms, chroma, augmentation, raga musicology basics)
• Setup guide (environment, downloading the Saraga subset, training your own run)

• The model classifies only the 10 ragas in the training subset; recordings in other ragas will be forced into one of the 10, and the report states this boundary.
• Near-neighbour ragas that share note sets (differing only in phrase movement) are the expected confusion cases — the confusion matrix documents them rather than hiding them.
• Performance depends on recording quality: heavy accompaniment, applause or poor microphones shift features away from the training distribution.
• The Saraga recordings are concert performances, so a model trained on them may not transfer to casual humming or harmonium renditions without retraining.
• No accuracy is claimed upfront: every reported number comes from the notebook's evaluation on your held-out recordings during your build.