Low Latency and Efficient Voice Activator for Edge Devices
Indian Space Research Organisation(ISRO) · Miscellaneous · Hardware
Crisp measurable metrics and a mature task make this winnable for a team with embedded skill — but it is judged on real hardware with limited novelty, so take it only if someone can do the TinyML and DSP work, and treat near-zero false activations as the hard target.
What it actually is
Voice devices that process everything in the cloud are slow, costly and privacy-invasive, so the edge should handle the wake-word and the cloud the rest. The ask is an ultra-light keyword-spotting model that runs on a low-power microcontroller — under 256KB RAM, under 10% CPU while idle — detects a custom wake word with near-zero false activations, and then streams the following audio to a cloud speech recogniser with minimal latency.
What to build
A TinyML keyword-spotting pipeline trained on a custom wake word — not a pre-trained 'Hey Google' — using only open-source frameworks like TensorFlow Lite for Microcontrollers, running within a hard budget of under 256KB RAM and under 10% CPU during continuous idle listening, achieving a high true-positive rate with near-zero false activations, and on detection immediately streaming the subsequent audio to a remote ASR server with minimal data overhead and minimal latency between keyword end and the cloud receiving the stream, delivered as a deployable architecture evaluated on real low-power microcontroller hardware.
Smallest thing that wins the room
On the actual microcontroller, say the custom wake word and show it triggering reliably while rejecting similar-sounding non-wake phrases, then show the subsequent audio reaching the ASR server with the measured latency and the idle RAM and CPU footprint within the stated limits.
How crowded this one gets
A guess, projected from the 2025 statements — the last year where both the submission counts and the winners were published.
Quieter than 87% of the 226 · #31 of 226 by expected field
Few teams are likely to go here. The best odds on the board come from statements like this.
Why: defence, intelligence and space bodies drew small fields; hardware halves the field a software statement gets.
This is a guess, not a fact
Nobody has published 2026’s numbers yet. This is an analysed estimate from last year’s pattern, so please do not take it as the truth — check the live counter on the SIH portal before you decide anything. The range covers the middle half of likely outcomes, so one statement in two lands outside it. Entry closes at 500 ideas per statement, so no range goes past that — a statement that reaches the cap fills and shuts rather than drawing an unlimited crowd. The model reads only three things a team can see before choosing — software or hardware, the theme, and what kind of body posted it — and those explain about a quarter of the variation in last year’s field sizes (R² 0.25 on held-out statements). Trust the band more than the number, and the ordering more than either. It cannot see how good your idea is, which is the part that actually decides it.
The scores
The number is the shorthand. The line under it is the reason.
Acceptance potential
3/5The metrics are crisp and objective which rewards a clean build, but this is a mature TinyML task with limited novelty, it demands genuine embedded and DSP skill, and physical-hardware evaluation favours teams with a hardware member over a software-led team looking for an easy target.
Feasibility
3/5Keyword spotting is a mature TinyML task with reference implementations, and the constraints are demanding but achievable with quantisation and a compact model, but it requires real embedded and audio DSP skill, hitting near-zero false activations is genuinely hard, and it is evaluated on physical hardware so a software-only team is at a disadvantage.
Innovation scope
2/5Keyword spotting under tight resource budgets is a well-solved TinyML problem with established architectures, so beyond the custom-keyword training and the streaming-latency optimisation there is little genuinely new to design.
Clarity
5/5The resource budget, the accuracy expectation, the latency metric, the open-source and custom-keyword restrictions and the physical-hardware evaluation are all stated precisely, making this one of the most measurable statements in the set.
Effort
HeavyTraining a compact KWS model, quantising and deploying it within a hard memory and CPU budget on real hardware, plus the low-latency streaming path, is focused embedded work.
Demo-ability
MediumA wake word triggering reliably on real hardware within the stated footprint is convincing, but it depends on having the microcontroller and a clean audio setup, and false activations in a noisy demo room are a risk.
In its favour
- Green flag: The metrics are crisp and objective — RAM, CPU, latency, false-activation rate — so success is measurable rather than argued
- Green flag: Keyword spotting is a mature TinyML task with reference architectures like DS-CNN to build on
- Green flag: The custom-keyword requirement is easy to satisfy by recording your own training samples
- Green flag: Open-source-only and no-pretrained-keyword rules level the field toward genuine engineering rather than SDK integration
Against it
- Red flag: Near-zero false activation during continuous listening is genuinely hard, and a demo room's noise is exactly where a KWS model produces false triggers
- Red flag: Evaluation is on physical low-power hardware, so a team without an embedded member is at a real disadvantage
- Red flag: The task is well-solved, so novelty is limited and the win comes from execution polish rather than a distinctive idea
- Red flag: Hitting the sub-256KB RAM budget while keeping accuracy high is a tight trade-off that punishes an unoptimised model
What you will be writing
- TensorFlow Lite for Microcontrollers
- Compact KWS (DS-CNN / tiny GRU)
- Int8 quantisation + model pruning
- MFCC / log-mel feature extraction on-device
- Low-latency audio streaming to ASR
- Deployment on <256KB-RAM microcontroller
- TinyML
- Keyword spotting
- Edge audio
Prior art to read before you start
custom wake-word detection · resource-constrained keyword spotting · low-latency edge-to-cloud audio
Analysed by Claude Opus. Every score above is a judgment call with its reasoning attached — kindly cross-check this against the official statement on the SIH portal before your team commits to it.