# Speech enhancement on a microcontroller · SIH 2026

By Aniruddha Kadam · Ongoing · since Aug 2026 · Updated 11 Oct 2026

Removing engine, gunfire and wind noise from radio speech on a $4 microcontroller, for Smart India Hackathon 2026: speech 11.2 dB clearer, measured on the board. Team of 6.

- **My part:** Capture firmware, on-device model, measurement rig (1 of 6)
- **Team:** 6 · Smart India Hackathon 2026, problem 52
- **Timeline:** Aug – Dec 2026
- **Status:** Ongoing
- **Links:** Firmware · Demo video · Raw measurements

## In short

- **Problem:** Speech on soldiers’ radios is buried in engine, gunfire and wind noise. The brief asks for enhancement on the device itself that reaches 3 targets: signal-to-noise ratio above 15 dB, an intelligibility score above 0.85 and a quality score above 2.5.
- **What I did:** Wrote the microphone capture firmware, fitted a compressed model onto the board and built the rig that measures it.
- **Result:** Speech 11.2 dB clearer (target: 15 dB), intelligibility 0.87, quality 2.6, on 120 held-out clips; 31 ms of delay at 38 mW. Clarity is still below target.
- **How to check:** Demo video · firmware · raw measurements (CSV) · reproduce with mic_lab.py

## Requirements: 3 audio targets, 128 KB of memory, under 40 ms
- Signal-to-noise ratio improved by more than 15 dB; intelligibility above 0.85; quality above 2.5
- Across steady, changing and sudden noise: engine, wind, gunfire
- Board: STM32F411, a 100 MHz Arm Cortex-M4F with 128 KB RAM and 512 KB flash, no external memory
- End-to-end delay under 40 ms, so a conversation still feels natural

> **Terms** **SNR**, signal-to-noise ratio: how much louder the speech is than the noise, in decibels. **STOI**: a 0–1 score for how intelligible speech is. **PESQ**: a score up to 4.5 for how good speech sounds to a listener.

## Block diagram: why an STM32F411, not an ESP32-S3
![Block diagram: INMP441 mic over I2S DMA into an STM32F411, model, DAC to headset, USB capture to mic_lab](https://anikadam.com/media/sih-block.svg)

_Signal path. The same mic and test tool also run on an ESP32-S3._

We tried an ESP32-S3 first. The STM32F411 won because its DMA timing is deterministic and CMSIS-NN runs the model faster on it.

## Measurements: 11.2 dB clearer, 31 ms, 38 mW on the board
_Video: 40-second demo, captioned: radio noise in, clean speech out, running on the board._

| Metric | Result | Target | How measured |
|---|---|---|---|
| Speech clarity (SNR gain) | +11.2 dB | > 15 dB | 120 held-out clips, 3 noise types |
| Intelligibility (STOI) | 0.87 | > 0.85 | pystoi, same 120 clips |
| Quality (PESQ) | 2.6 | > 2.5 | pesq, same 120 clips |
| End-to-end delay | 31 ms | < 40 ms | click track in vs out on a 2-channel scope |
| Power | 38 mW | — | INA219 on the 3.3 V rail, 60 s average |
| Flash / RAM | 214 / 97 KB | 512 / 128 KB | arm-none-eabi-size |

![Spectrogram of noisy input and enhanced output](https://anikadam.com/media/sih-spectro.svg)

_Before and after, sudden (impulsive) noise._

## Debugging: why the samples were off by one bit
**Symptom:** the captured audio was loud crackle; the values looked right but shifted. **Hypotheses:** wrong sample rate, wrong word length, wrong I²S standard. **Measurement:** a logic-analyser capture of the clock and data lines.

![Logic analyser capture of I2S clock and data](https://anikadam.com/media/sih-scope.svg)

_Clock and data at the microphone._

**Root cause and fix:** the INMP441 sends 24-bit data one clock after the word-select line changes (the Philips standard), but the peripheral was set to MSB-justified. Switching the standard fixed it.

## What broke: sudden noise like gunfire improves least
- **What failed:** Sudden noise like gunfire improves by only 6.1 dB, against 11.2 dB overall.
- **How I found out:** Splitting the 120 test clips by noise type.
- **What I changed:** Moving the spectral transform to CMSIS-DSP to free 18 KB of RAM for a larger model.
- **Still open:** Whether the larger model fits inside the 40 ms delay budget.

## References

1. [MLPerf Tiny benchmark](https://arxiv.org/abs/2106.07597)
2. [CMSIS-NN](https://github.com/ARM-software/CMSIS-NN)

**Credits.** Arya Maddel, Jagdish Sonigra, Kashish Degaonkar, Parth Latke and Tanish Thakkar.

**How I used AI tools.** Used Claude to review the DMA code and draft this write-up; all measurements taken by hand.

Source: https://anikadam.com/work/sih26/