Cascade Hotword Detection for Noise-Robust Low-Power Wake Words
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing voice-enabled devices face challenges in efficiently detecting hotwords with noise robustness while balancing power consumption, latency, and computational load, especially on battery-powered devices with limited resources.
Innovation Solution
A cascade hotword detection architecture is employed, where a first-stage hotword detector on a digital signal processor coarsely screens multi-channel audio for candidates, triggering a second-stage detector on an application processor for accurate verification, leveraging noise cleaning algorithms to enhance detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a single-stage hotword detector processes all audio data, then detection accuracy can be maintained, but power consumption and computational load increase significantly
Solution Approach 1:
The hotword detection system is segmented into two stages: a first-stage detector that performs coarse screening on all audio data with lower computational requirements, and a second-stage detector that performs fine verification only on candidate segments identified by the first stage. This segmentation allows the system to maintain high detection accuracy while significantly reducing overall power consumption by keeping the computationally intensive second-stage detector inactive most of the time.
2Reliability
If noise cleaning algorithms are applied to all audio data, then noise robustness improves, but processing time and computational load increase
Solution Approach 1:
The system applies noise cleaning algorithms as a preliminary action only to audio segments that are identified as hotword candidates by the first-stage detector. By performing noise cleaning selectively on candidate segments rather than on all incoming audio data, the system achieves improved noise robustness for critical detections while minimizing the overall processing time and computational load.
3Measurement precision
If a second-stage detector is added for verification, then false acceptance rate decreases, but system complexity increases
Solution Approach 1:
The detection system is divided into two functional segments: a first-stage detector that quickly identifies potential hotword candidates and a second-stage detector that performs thorough verification. This segmentation reduces false acceptance rates by introducing a verification layer, while the modular design keeps system complexity manageable through clear separation of concerns and optimized resource allocation.
4Reliability
If multi-channel audio is processed, then noise robustness improves, but computational load and power consumption increase
Solution Approach 1:
The system processes multi-channel audio by segmenting the processing pipeline: the first-stage detector performs coarse screening on multi-channel input to identify candidate segments, and only these candidates are passed to the second-stage detector for fine verification with full noise cleaning. This segmentation enables the system to leverage the noise robustness benefits of multi-channel processing while significantly reducing power consumption by limiting intensive computations to small candidate segments rather than continuous multi-channel audio streams.
Data Source
Figure 1
Figure 2A
Figure 2B
AI summary
A method (400) includes receiving, at a first processor (110) of a user device (102), streaming multi-channel audio (118) captured by an array of microphones (107), each channel (119) including respective audio features. For each channel, the method also includes processing, by the first processor, using a first stage hotword detector (210), the respective audio features to determine whether a hotword is detected. When the first stage hotword detector detects the hotword, the method also includes the first processor providing chomped raw audio data (212) to a second processor that processes, using a first noise cleaning algorithm (250), the chomped raw audio data to generate a clean monophonic audio chomp (260). The method also includes processing, by the second processor using a second stage hotword detector (220), the clean monophonic audio chomp to detect the hotword.