Wake-Word Detection

Wake-word engines must run continuously without draining the battery. This demands microwatt-level power budgets achieved through dedicated audio DSPs or ultra-low-power microcontrollers.

• Typical power consumption: < 1 mW for the always-on frontend
• Uses hardware-based audio activity detection to gate the neural network
• Only activates the larger inference model when acoustic energy exceeds a threshold
• Enables months-to-years of battery life on coin-cell powered medical wearables

To balance accuracy with power efficiency, wake-word systems employ a cascaded detection pipeline. A lightweight first stage filters obvious non-speech, while a more accurate second stage confirms the trigger phrase.

• Stage 1 (Voice Activity Detection): Simple energy-based or lightweight DSP algorithm that rejects silence and noise with near-zero power
• Stage 2 (Keyword Spotting): A small neural network (typically 100-500 KB) processes only the audio frames that pass Stage 1
• This cascade reduces the duty cycle of the neural network by 90-99% in quiet environments
• Critical for medical voice assistants that must respond reliably in clinical settings

Unlike batch inference on recorded audio, wake-word detection operates on a continuous streaming input. The model processes overlapping audio frames in real-time to minimize detection latency.

• Audio is segmented into overlapping windows (typically 25-40 ms) with a stride of 10 ms
• Feature extraction converts raw PCM audio to Mel-frequency cepstral coefficients (MFCCs) or log-mel filterbank energies
• The neural network maintains a sliding context window of recent frames to capture the temporal structure of the trigger phrase
• Streaming inference ensures sub-100 ms detection latency for responsive voice-controlled medical devices

Wake-word models must distinguish the target phrase from phonetically similar utterances with extremely low false accept rates. This requires discriminative training on diverse negative examples.

• Models are trained on the target phrase plus thousands of hours of negative speech and background noise
• Hard negative mining identifies phonetically confusable phrases (e.g., 'Hey Siri' vs. 'Hey Sarah')
• Modern architectures use temporal convolutions or attention mechanisms to model the sequential phoneme structure
• Typical false reject rate: < 1% at one false alarm per hour in quiet conditions

Deploying wake-word models on medical microcontrollers requires aggressive compression techniques that maintain accuracy within tight memory and compute constraints.

• Post-training int8 quantization reduces model size by 4x with minimal accuracy loss
• Structured pruning removes entire channels to create hardware-friendly sparse models
• Knowledge distillation trains a compact student model (50-200 KB) from a larger teacher
• Operator fusion combines convolution, batch norm, and activation into single kernels
• Typical footprint: 100-500 KB for the neural network, fitting within the SRAM of a Cortex-M4