Dual Neural Network ASR for Low-Power Continuous Speech Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing continuous automatic speech recognition (ASR) systems face challenges in minimizing power consumption while recognizing a limited vocabulary in continuous audio streams, as they often process unwanted words and sounds, leading to unnecessary power usage and require continuous operation without user prompting.
Innovation Solution
The implementation of a dual neural network architecture, where a state-maintaining neural network controls a non-state-maintaining neural network to perform speech recognition only when sufficient feature vectors representing a recognizable word are available, reducing unnecessary processing and power consumption by allowing the system to operate in a low-power quiescent state during non-speech activity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If continuous speech recognition processing is performed on all audio data, then speech recognition accuracy is improved, but power consumption increases
Solution Approach 1:
The system segments the neural network processing into two distinct parts: a state-maintaining neural network that continuously monitors audio data in low-power mode, and a non-state-maintaining neural network that performs full speech recognition processing only when triggered. This segmentation allows the system to maintain speech recognition accuracy while significantly reducing power consumption during non-speech periods.
Solution Approach 2:
The system implements periodic action by having the state-maintaining neural network continuously monitor audio data and trigger the non-state-maintaining neural network only when speech is detected. This periodic activation pattern ensures that full processing power is applied only when necessary, reducing overall power consumption while maintaining recognition accuracy.
2Reliability
If the system processes all audio data continuously, then speech recognition reliability is improved, but unnecessary power usage increases
Solution Approach 1:
The state-maintaining neural network performs preliminary action by continuously monitoring audio data in a low-power state and identifying potential speech segments. This preliminary filtering ensures that the more power-intensive non-state-maintaining neural network is activated only when speech is actually present, eliminating unnecessary power usage while maintaining recognition reliability.
Solution Approach 2:
The state-maintaining neural network acts as an intermediary between the continuous audio input and the non-state-maintaining neural network. It filters and prepares audio data, triggering the main processing network only when speech is detected, thus reducing unnecessary power consumption while ensuring reliable speech recognition.
3Use of energy by moving object
If the system operates in low-power mode, then power consumption is reduced, but the ability to recognize speech in real-time may be compromised
Solution Approach 1:
The system dynamically adjusts its operational state based on the presence of speech. The state-maintaining neural network continuously monitors audio input and dynamically triggers the non-state-maintaining neural network when speech is detected. This dynamic switching ensures real-time speech recognition capability while maintaining low power consumption during non-speech periods.
Solution Approach 2:
The state-maintaining neural network provides continuous feedback about the audio input state to control activation of the non-state-maintaining neural network. This feedback mechanism ensures that the system responds immediately to speech detection, maintaining real-time recognition capability while optimizing power consumption based on actual speech presence.
Data Source
AI summary
Automatic speech recognition (ASR) systems and methods. An ASR system includes a vector memory to store a plurality of feature vectors sequentially extracted from an audio data stream. A first neural network performs speech recognition processing on feature vectors stored in the vector memory to attempt to recognize a word from a predetermined vocabulary. A second neural network controls when the first neural network performs the speech recognition processing. The first neural network is held in a quiescent state except when performing the speech recognition processing under control of the second neural network.