Acoustic Trigger Detection Using Time-Delay Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Speech recognition systems face challenges in devices with limited computational capacity, leading to high error rates in trigger detection and excessive resource consumption, particularly due to the need for continuous audio data processing and potential false alarms.
Innovation Solution
A spoken language processing system that employs an acoustic front end, feature extractor, and Time-Delay Neural Network (TDNN) for efficient trigger detection, limiting audio data transmission to only detected trigger intervals, and using a neural network structure with time delays and weight decomposition to reduce computational requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If speech recognition is performed continuously to detect trigger words, then trigger detection accuracy is improved, but computational resource consumption increases
Solution Approach 1:
The system segments the audio processing task into two distinct stages: (1) a lightweight trigger detection stage that continuously monitors for wake words using minimal computational resources, and (2) a full speech recognition stage that is only activated when a trigger is detected. This segmentation allows the system to maintain high trigger detection accuracy while avoiding continuous heavy computational processing, thus resolving the contradiction between detection accuracy and resource consumption.
Solution Approach 2:
The system performs preliminary trigger detection using a simplified acoustic model before committing to full speech recognition processing. This preliminary action filters out non-trigger audio segments, ensuring that computationally intensive speech recognition is only performed when necessary. The preliminary trigger detection maintains accuracy by using dedicated acoustic features while preventing wasteful resource consumption on non-trigger segments.
2Measurement precision
If audio data is transmitted continuously to the server, then speech recognition accuracy is improved, but network bandwidth consumption increases
Solution Approach 1:
The system extracts and transmits only the relevant audio segments containing detected trigger words to the server, rather than transmitting continuous audio streams. This extraction approach ensures that the server receives sufficient audio data for accurate speech recognition while dramatically reducing network bandwidth consumption by eliminating transmission of non-trigger audio segments.
3Reliability
If continuous audio processing is performed, then trigger detection reliability is improved, but false alarm rate increases
Solution Approach 1:
The system applies partial action by using a simplified trigger detection model that processes audio continuously with minimal resources, rather than applying full speech recognition continuously. This partial processing approach maintains sufficient reliability for trigger detection while avoiding the excessive sensitivity that would cause false alarms, as the simplified model is specifically tuned for wake word detection rather than general speech recognition.
Data Source
AI summary
A method for selective transmission of audio data to a speech processing server uses detection of an acoustic trigger in the audio data in determining the data to transmit. Detection of the acoustic trigger makes use of an efficient computation approach that reduces the amount of run-time computation required, or equivalently improves accuracy for a given amount of computation, by using a neural network to determine an indicator of presence of the acoustic trigger. In some example, the neural network combines a “time delay” structure in which intermediate results of computations are reused at various time delays, thereby avoiding computation of computing new results, and decomposition of certain transformations to require fewer arithmetic operations without sacrificing significant performance.