That’s when I had an idea: What if I could build an AI-powered tool that could recognize Hindustani classical ragas from any given audio? Something like Shazam, but for classical Indian music.

The Spark That Led to Raga-Match

I started researching how music recognition works, digging into the underlying technology behind apps like Shazam. The key to their success was something called spectrogram-based audio fingerprinting, where an audio file is broken down into its frequency components over time, creating a visual representation of sound.

But Hindustani classical music is very different from mainstream Western music. Unlike fixed-scale compositions, raagas are fluid, with intricate ornamentations (meend, gamak) and variations in tempo (layakari). This made the challenge even more exciting: I couldn’t just replicate Shazam’s algorithm; I had to tailor it specifically for classical music.

The Data Problem: No Dataset, No Problem?

Unlike Western music, which has well-documented datasets, Hindustani classical music doesn’t have a structured, easily accessible dataset for raga recognition. This was my first major roadblock. I had to manually collect audio samples, look up recording snippets, and organize them all into a single dataset.

At first, it seemed like a simple task—until I started encountering audio encoding errors while processing the files. Different recording formats, metadata issues, and even missing audio links forced me to update file paths multiple times. But after weeks of trial and error, I finally compiled a working dataset, which I later uploaded to Kaggle. You can check it by following the link below:
🎵Hindustani Classical Music Raga Dataset

The First Roadblock: Dealing with Noisy Audio

One of the first problems I encountered was the quality of audio recordings. Unlike studio-produced songs, many Hindustani classical performances are recorded live, sometimes with background noise or accompanying instruments overpowering the vocals.

To tackle this, I integrated noise reduction techniques using Librosa and SciPy. I applied:

✔ Preprocessing filters to remove background noise

✔ Normalization techniques to balance volume levels

✔ Spectrogram generation to highlight important frequency patterns

Building the Model: CNN + LSTMs for Raga Identification

Once I had the dataset, it was time to train the AI model. Hindustani classical music is vastly different from Western music because:
Ragas evolve over time, unlike pop songs with fixed chord progressions.
Ornamentations like meend, gamak, and tanpura drones make feature extraction difficult.
There are no "exact" labels for ragas, as performances often mix elements of multiple ragas.

Instead of using traditional audio fingerprinting, I decided to build a deep learning model that could analyze Mel-Spectrograms and MFCCs (Mel-Frequency Cepstral Coefficients). These are visual representations of sound, making them perfect for a Convolutional Neural Network (CNN) + LSTM (Long Short-Term Memory) model.

• CNN layers helped extract spatial features from the spectrograms.

• LSTMs captured temporal dependencies in how a raga unfolds over time.

After multiple training iterations, the model performed shockingly well! The .pth file containing the trained model gave results that were too good to be true. But, AI models performing well in theory doesn’t mean they’ll work perfectly in real-world applications—the real challenge was yet to come.

Real-Time Identification: The Web App Challenge

Once the model started making decent predictions, the next step was to build an interactive system. I envisioned a web app with two features:
🎵 Upload an audio file → The app processes it and returns the raga name.
🎙 Record live singing → The app analyzes the real-time input and classifies the raga.

To achieve this, I built the backend using FastAPI, a high-performance web framework that supports Web Sockets for real-time updates. The biggest challenge? Live audio streaming and processing. Unlike static files, a live recording had to be broken into chunks, processed on the go, and classified in real time without delays. This required fine-tuning buffer sizes and optimizing the Fast Fourier Transform (FFT) calculations.

The Flask API Nightmare and WebSockets Struggles

The next step was to integrate the model into a Flask API so it could be accessed via a web application. I planned a Next.js frontend where users could:
✅ Upload an audio file to get the raga name.
✅ Record live singing, and the AI would identify the raga in real time.

I expected this to be straightforward, but I underestimated the difficulties of handling real-time audio processing with WebSockets. Flask wasn’t cooperating, and setting up WebSockets for live audio streaming became a major headache.

Real-time latency issues made it hard to analyze continuous audio input.
Buffering problems caused unexpected delays in raga identification.
Deployment challenges meant Flask and Next.js weren’t communicating properly.

Still a Work in Progress... But Closer Than Ever!

Even though I faced several roadblocks, Raga-Match is finally taking shape. The AI model works well, and I’ve successfully preprocessed the dataset, trained the CNN-LSTM model, and integrated the API—but live audio analysis is still a work in progress.

I’m currently experimenting with FastAPI instead of Flask, optimizing WebSockets, and fine-tuning real-time inference. The ultimate goal is to deploy this as a fully functional AI-powered web app, making Hindustani classical music recognition accessible to everyone.

Final Thoughts: The Beauty of Music and AI

Building Raga-Match has been an incredible experience. It started as a simple idea—combining my love for music and coding—but turned into a massive learning journey in AI, data collection, and web app deployment.

Despite all the struggles, seeing the model correctly classify ragas feels like magic. I still have a long way to go, but one thing’s for sure: Music and AI belong together, and this is just the beginning!

When I first delved into the neuroscience of DID, I was struck by how little we understood about its underlying mechanisms. DID is often dismissed as purely psychological, yet mounting evidence suggests that abnormal neural connectivity and plasticity play a crucial role. This realization pushed me to explore the disorder from a computational and neurobiological perspective, integrating machine learning with functional neuroimaging to uncover patterns that traditional approaches might overlook.

Mapping the Neural Circuit Dysfunction in DID

One of the most challenging aspects of studying DID is the heterogeneity in how it manifests. Individuals with the disorder experience disruptions in memory, sense of self, and identity switching, which are thought to arise from dysregulated interactions between key brain regions. My research focused on analyzing resting-state functional MRI (rs-fMRI) data using Convolutional Neural Networks (CNNs) with attention mechanisms and saliency maps to identify the most affected brain regions.

Key Brain Regions Implicated in DID

Through deep learning models, I examined the connectivity between:

• Ventromedial prefrontal cortex (VmPFC) – Crucial for self-referential processing and identity integration. Reduced connectivity here is associated with self-fragmentation and altered self-awareness in DID.

• Amygdala – Governs emotional regulation, often hyperactive in trauma-related disorders, contributing to fear responses and dissociation triggers.

• Hippocampus – Essential for memory formation, frequently impaired in DID, leading to memory gaps and disrupted autobiographical recall.

• Insular Cortex – Plays a role in interoception (awareness of bodily states), with abnormalities linked to dissociative symptoms.

• Posterior Parietal Cortex (PPC) – Involved in spatial and sensory integration, often showing reduced connectivity in DID, contributing to feelings of detachment and depersonalization.

To enhance interpretability, I incorporated attention mechanisms with CNNs and saliency maps to pinpoint which brain regions exhibited the most abnormal connectivity patterns in resting-state fMRI. The attention mechanisms helped the model prioritize clinically relevant neural signals, ensuring that the most diagnostically significant regions were highlighted. Saliency maps provided a visual explanation of neural disruptions, allowing us to see precisely where DID-related abnormalities occurred. This approach not only improved classification accuracy but also offered valuable insights into the functional deficits underlying DID.

By leveraging these techniques, I was able to observe distinct patterns of hypo- and hyperconnectivity across these regions, supporting the hypothesis that DID arises from impaired communication between self-referential and emotion-processing networks. These findings suggest that clinicians might detect DID in patients by focusing on abnormalities in identity processing, emotional regulation, and interoception, as governed by these affected regions. Future advancements in this field could lead to the development of personalized diagnostic tools based on neural biomarkers, paving the way for targeted interventions tailored to each individual’s unique neural profile.

Bridging Computational Neuroscience with Clinical Insights

One of the most rewarding aspects of this work has been its potential clinical applications. Many DID patients struggle with misdiagnosis and inadequate treatment due to the subjective nature of psychiatric evaluations. By leveraging functional connectivity biomarkers, we can move toward more objective diagnostic tools, ensuring that individuals receive the care they need.

Beyond diagnosis, my work has also explored predictive modeling for therapeutic interventions. By studying long-term potentiation (LTP) and long-term depression (LTD) in synaptic plasticity, I aimed to uncover whether targeted neuroplasticity-based treatments—such as non-invasive brain stimulation or AI-personalized cognitive therapy—could help reintegrate fragmented neural networks in DID.

The Future of DID Research and My Next Steps

While my findings have contributed a small piece to the larger puzzle of DID, there is still much to explore. Moving forward, I aim to integrate calcium-dependent synaptic plasticity modeling to better understand how neural circuits in DID patients adapt (or fail to adapt) over time. Computational psychiatry offers a revolutionary approach to mental health, and I believe that merging neuroscience with AI and deep learning will be instrumental in unraveling the complexities of DID.

This journey has reinforced my belief that identity is not just a psychological construct but a deeply biological phenomenon. The brain’s ability to fragment and reorganize in response to trauma is both awe-inspiring and deeply tragic. As I continue my research, my ultimate goal is to bridge the gap between neuroscience and psychiatry, using computational models to illuminate the hidden architecture of the mind—and, hopefully, to offer new paths toward healing.