Dual Pipeline Wake-Up Phrase Detection Architecture
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The implementation of wake-up phrase detection in speech recognition devices leads to increased power consumption and latency due to the continuous operation of microphone and speech detection circuitry, which remains in a powered state for extended periods, affecting the responsiveness and efficiency of speech recognition tasks.
Innovation Solution
A dual pipeline approach is implemented, where a high latency and low audio quality pipeline is used for speech onset detection, and a low latency and high audio quality pipeline is used for wake-up phrase detection, allowing the speech onset detection circuitry to operate at a lower power consumption and clock rate, while the wake-up phrase detection circuitry operates at a higher power consumption only after a speech onset event is detected, utilizing a history buffer to minimize latency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If wake-up phrase detection is implemented, then speech recognition capability is improved, but power consumption increases
Solution Approach 1:
The system divides wake-up detection into two separate pipelines: a low-power speech onset detection pipeline that operates continuously, and a high-performance wake-up phrase detection pipeline that activates only when speech is detected. This segmentation allows the system to maintain speech recognition capability while minimizing power consumption during idle periods.
Solution Approach 2:
The system dynamically transitions between power states based on speech activity. The speech onset detector triggers a state change from low-power mode to high-performance mode when speech is detected, allowing the wake-up phrase detection circuitry to operate at full performance only when needed.
2Measurement precision
If wake-up phrase detection circuitry operates continuously, then detection accuracy is improved, but latency increases
Solution Approach 1:
The speech onset detection pipeline operates continuously in advance to detect the beginning of speech. When speech onset is detected, it triggers the wake-up phrase detection pipeline, allowing the system to prepare for and respond to wake-up phrases more quickly without requiring continuous operation of the full detection system.
Solution Approach 2:
The speech onset detector acts as an intermediary between the continuous audio input and the high-performance wake-up phrase detection system. It filters and pre-processes the audio stream, triggering the full detection pipeline only when relevant speech activity is detected, thereby reducing overall latency.
3Measurement precision
If high quality audio processing is used for speech onset detection, then detection accuracy is improved, but power consumption increases
Solution Approach 1:
The system applies different audio quality processing to different detection tasks. The speech onset detection pipeline uses lower quality audio processing sufficient for detecting speech beginnings, while the wake-up phrase detection pipeline uses high quality processing only when needed. This local differentiation of quality levels optimizes power consumption while maintaining necessary detection accuracy.
Solution Approach 2:
The system changes audio processing parameters based on the detection stage. The speech onset detector operates with reduced audio quality parameters (lower sampling rate, reduced processing depth), while the wake-up phrase detector uses full-quality parameters. This parameter adaptation allows accurate detection at each stage while minimizing overall power consumption.
Data Source
AI summary
A phrase detection device includes a high latency pipeline to transmit a first portion of audio data from an audio data source to a processing unit, where the high latency pipeline includes a history buffer to store the first portion of the audio data, and a low latency pipeline to transmit a second portion of the audio data from the audio data source to the processing unit with a lower latency than the high latency pipeline. A sound onset detector coupled with the audio data source detects a sound onset event based on the audio data. A synchronization circuit coupled with the high latency pipeline and the low latency pipeline, in response to the sound onset event, synchronizes output to the processing unit of the first portion of the audio data stored in the history buffer and the second portion of the audio data via the low latency pipeline.