Low Latency and Efficient Voice Activator for Edge Devices
“We are building an ultra-efficient voice assistant trigger that runs on tiny microcontrollers without needing the internet. When it hears a specific secret command word, it instantly wakes up and securely streams subsequent voice instructions to a remote cloud server.”
Deconstructed Problem Blueprint
WHAT TO BUILD?
An embedded TinyML Keyword Spotting (KWS) firmware paired with a low-latency audio streaming daemon running on an edge microcontroller (like an ESP32 or Raspberry Pi Pico W).
WHO IS IT FOR?
ISRO space scientists, field engineers, and mission control operators who require hands-free, low-power, offline voice activation for equipment and IoT setups.
WHY DOES IT EXIST?
Relying entirely on cloud processing for voice assistants causes high latency, severe privacy risks, and excessive battery/bandwidth consumption, making edge wake-word detection essential.
INPUT → PROCESS → OUTPUT ARCHITECTURE PIPELINE
Continuous real-time audio feed from an I2S digital microphone (e.g., INMP441) capturing ambient sound.
Audio framing, Mel-Frequency Cepstral Coefficients (MFCC) feature extraction, inference via a quantized TinyML model (<256KB RAM footprint), followed by packetization and UDP/WebSocket streaming upon keyword confirmation.
An immediate hardware interrupt trigger and a low-latency compressed audio stream directed to a remote Automated Speech Recognition (ASR) server.
CONSTRAINTS
- Strict memory footprint: Under 256KB of RAM available on the target microcontroller.
- Low CPU usage: Under 10% utilization during idle continuous listening mode.
- No proprietary SDKs or pre-trained generic wake words ('Hey Google', 'Alexa'); must train on a custom keyword.
- Open-source frameworks only (TensorFlow Lite for Microcontrollers or PyTorch Mobile).
JURY METRICS
- Model RAM and Flash memory footprint
- Idle CPU utilization percentage
- True-positive keyword detection accuracy with near-zero false-positive rates
- Latency delta between keyword ending and cloud ASR reception
CORE BOTTLENECK
Balancing high keyword detection accuracy with extremely strict RAM (<256KB) and low idle CPU (<10%) limits on resource-constrained microcontrollers without triggering false positives from background noise.
Why You Should Choose This PS (Pros)
- Extremely clear problem scope with precise, measurable evaluation metrics.
- Hardware track solutions with live physical demos always stand out to judges.
- Direct relevance to space tech automation, IoT, and defense-grade edge computing.
Challenges & Risks to Watch (Cons)
- Requires actual hardware setup (ESP32/Raspberry Pi + microphone) during presentation, which can fail unexpectedly.
- Training and quantizing custom acoustic models from scratch requires dataset collection and signal processing expertise.
Recommended Tech Stack & Grand Finale Features
Killer Features to Impress the SIH Jury
View Official SIH 2026 Portal Raw Text▼
Background As voice-controlled IoT proliferate, processing everything in the cloud is too costly, privacy-invasive, and slow. The future belongs to hybrid architectures where the edge handles the initial 'wake-up' and the cloud handles the heavy lifting.
Description Build an ultra-lightweight, highly accurate keyword spotting (KWS) model that runs locally on a low-power device. Upon detecting the keyword, the system must instantly and efficiently stream the subsequent audio to a remote Automated Speech Recognition (ASR) server with minimal data overhead and latency.
Key Metrics for Evaluation
Software & Framework Restrictions
Expected Solution Teams are expected to deliver a robust, deployable system architecture. A successful submission must strictly satisfy the following technical boundaries: