Real-Time Audio DSP · Acoustics · Embedded Systems/Firmware
_{Seven years shipping audio firmware on constrained hardware — now working on the research questions underneath it.}

Adaptive feedback and howling suppression in closed acoustic loops — detection, notch allocation, frequency shifting, and how much stable gain you can actually buy. Sound-field reconstruction with calibrated uncertainty — when a spatial audio system says "the field here is X", what guarantee comes with that, and when should it refuse to answer? Real-time DSP on constrained hardware — what survives the trip from a MATLAB prototype to a DSP core with a fixed cycle budget, and what quietly doesn't. Physical and perceptual modelling for spatial and immersive audio.

The through-line: I've spent my career on the side of audio where the algorithm has to run — in a car cabin, a hospital corridor, a smart ring. I'm now interested in the part where it also has to be provably right.

Both are self-contained, documented, and reproducible from a clean checkout.

An open-source FAUST implementation of in-car communication (ICC): cabin speech picked up, processed, and replayed through the same cabin's loudspeakers — a closed electro-acoustic loop that wants to howl.

• Filterbank howling detector rather than FFT — FAUST expands FFT into scalar butterflies, so the filterbank is the structure that actually stays real-time
• Self-allocating notch cascade — slots claimed and released as howling frequencies come and go
• Frequency shifting via fi.pospass for additional loop-gain margin
• Measured, not just compiled: +4 dB MSG bypassed vs +16 dB suppressed — 12 dB of added stable gain on the reference synthetic cabin at 16 kHz, recorded in docs/tuning.md as a regression check
• Two bugs surfaced only by running it: an unnormalised cabin path with +9 dB peak gain (loop-gain readings were physically meaningless until fixed) and unstable loops reaching NaN, now soft-saturated

lib/icc.lib + three designs · ~700 lines of FAUST · verified against FAUST 2.70.3 · C/C++/Rust backends confirmed · CI · CITATION.cff

Conformal prediction gives distribution-free coverage, but standard weighted methods rely on estimated likelihood ratios $dQ/dP$, where estimation errors degrade coverage. In sound-field reconstruction, $dQ/dP$ is known analytically from spatial geometry, enabling exact coverage guarantees without estimation step errors. It replaces density-ratio estimation with analytical spatial ratios to achieve 0.901 coverage vs. 0.900 nominal (compared to 0.765 for split conformal), while explicitly abstaining (~87% infinite prediction intervals) outside calibration support instead of failing silently (0.283 coverage). Furthermore, it proves that effective sample size is maximized when the calibration microphone layout directly mirrors the target query density, providing an optimal physical placement strategy in a pure base MATLAB implementation with reproducible synthetic image-source generation.

Full-duplex dsp chain design for 1mic voice communication chain on AudioWeaver

A MetricGAN-inspired framework for acoustic feedback suppression that replaces non-differentiable loop bifurcation points with a learned surrogate predictor (StabilityNet) to directly optimize notch filter allocation. It is an extension of In-CarCommunication or Faust-ICC

• Differentiable MSG Estimation: Trains a neural surrogate to estimate continuous Maximum Stable Gain (MSG) headroom, enabling end-to-end policy optimization where physical loop oscillation prevents direct gradient backpropagation.
• Group Delay & Phase Shift Dynamics: Quantifies how notch bank group delay alters loop phase alignment. High-order filters (Length 513) destabilize the loop (−8.00 dB stable gain) by shifting howling to adjacent frequencies, whereas lower-order filters (Length 31) achieve +7.75 dB of added stable gain.
• Oracle-Placed Benchmarks: Isolates physical phase delay from detection errors by testing against ground-truth howling frequencies, proving filter length trade-offs are physical rather than algorithmic.

Blind SRO estimation for asynchronous air-microphone and body-conduction sensor fusion. It introduces class-conditioned coherence drift to handle signal-dependent acoustic channels, includes reproducible Python experiments, unsupervised conditioning studies, and a FAUST real-time DSP kernel. This is validated through experiments, with real-speech validation planned using the Vibravox dataset.

A multi-modal audio-visual tracking framework leveraging Random Finite Set (RFS) filter dynamics for dynamic spatial speaker localization and state estimation under visibility shifts.

• Random Finite Set Formulation: Handles varying talker counts, track creation, and track loss without explicit target-to-observation data association overhead or track-switching failures under multi-speaker overlap.
• Audio-Visual Sensor Fusion: Fuses spatial audio Direction-of-Arrival (DoA) vectors with visual detection bounding boxes to maintain continuous trajectory estimation during severe visual occlusions.
• Validation Scope & Upper Bounds: Evaluated under measurement-level baseline sweeps to establish theoretical upper bounds across visibility regimes prior to full end-to-end evaluation on real-world datasets like AVA-AVD.

A high-throughput computer vision pipeline designed for spatial target detection, temporal trajectory estimation, and robust feature extraction under dynamic lighting and severe occlusions.

• Temporal Feature Association: Implements multi-frame feature tracking and temporal bounding box association to eliminate track fragmentation during transient target occlusions.
• Low-Latency Spatial Inference: Optimizes visual backbone architectures (TorchScript/TensorRT) to achieve real-time frame processing within strict edge-compute latency budgets ($< 15\text{ ms}$).
• Robustness & Domain Generalization: Evaluated against challenging spatial datasets featuring variable illumination, dynamic camera movement, and dense multi-object interaction.

Simulation modules and scripts for Delay-Tolerant Networking (DTN) in Low Earth Orbit (LEO) constellations using ns-3.

• ns-3 scenario scripts modelling intermittent contacts, store-and-forward routing, and contact plan generation for LEO satellites
• Bundle protocol and routing evaluation (latency, delivery ratio, buffer utilisation) across constellation and ground-station topologies
• Reproducible experiments with run scripts, parameter sweeps, and plotting utilities; documentation in README with quickstart and dependencies

Seven years of production audio and systems firmware — the reason I care about cycle budgets and failure modes rather than just algorithms.

• IEEE 2026 — Author-oriented semantic plagiarism detection using transformer architectures
• IEEE 2026 — Transformer-based authorship attribution: fine-tuning BERT and S-BERT for stylometric analysis
• arXiv 2025 — Conversational AI dialog for medicare, powered by fine-tuning and retrieval-augmented generation
• 2024 — AI-engine-based acceleration for high-performance programmable SoC designs
• 3 published Indian patents (Dec 2024) — semantic role labelling for sentiment analysis · RAG-based conversational AI · fine-tuned BERT for relation extraction and NER
• In preparation — neuromorphic drowsiness prediction with multi-sensor fusion

🔗 ORCID 0009-0001-5506-9588

Audio & DSP — FAUST · MATLAB · Audio Weaver · Qualcomm AudioReach · Android Audio Framework · ALSA · Kalimba DSP · IIR/FIR design, AEC, AGC, noise suppression

Systems — C · C++ · Python · Zephyr · FreeRTOS · ThreadX · Linux kernel · ARM Cortex-M3/M7/A53 · x86 · BLE, I2S, I2C, UART · JTAG/J-Link/Trace32 · Git, Gerrit, CI

Graduated in keyboard from Bangiya Sangeet Parishad affiliated with Rabindra Bharati University (Kolkata) and recognized by Indira Kala Sangeet Vishwavidyalaya and the Government of Assam, and graduated in painting from Tripura Fine Arts Academy with first division and distinction. Dhwaani — my handle — means Resonance.

Happy to talk about adaptive feedback control, spatial audio, uncertainty quantification, or getting DSP to survive on a small core.