Record 05
Dhwani-Kavach
Live dual-microphone speech enhancement for noisy environments, built for Smart India Hackathon 2026 (SIH26052, DRDO).
- Status
- Bench-tested
- Year
- 2026
- Built on
- FastEnhancer (ICASSP 2026)
- ΔPESQ vs offline
- 0.0363 (target < 0.05)
- Latency
- not yet measured
Briefing
- Problem
- Keep speech usable in rotor and machinery noise, live.
- What exists today
- A real-time dual-microphone pipeline, bench-tested, built on FastEnhancer.
- Key result
- The real-time path stays within 0.05 PESQ of the offline model (ΔPESQ 0.0363). Latency is not yet measured.
Objective
Keep speech usable in rotor and machinery noise, live. The real-time path must match the same model run offline, and switching between raw and enhanced audio must not click.
Architecture
Schematic
-
Input
- Primary microphone
- Reference microphone
- Link: 48 kHz
Conditioning
- 48 kHz → 16 kHz
- High-pass and notch filters
- Link: 16 kHz
Cancellation
- NLMS on the reference
- FastEnhancer (ONNX)
-
Output
- Suppression floor, AGC, limiter
The neural stage is FastEnhancer, a published streaming speech-enhancement model. My work is the live chain around it, the operator panel, and the acceptance tests.
Measurements
Acceptance tests and the demo.
- ΔPESQ 0.0363 Real-time path within 0.05 PESQ of offline: pass. Acceptance test
- 1.088 → 1.602 PESQ, noisy to enhanced, on one synthetic demo sample. Synthetic sample
- — End-to-end loopback latency. Not yet measured
Status
Bench-tested
Links
End of record R-05
Questions about this record? hello@vedantsomani.tech or open a channel.