Neural Network Waveform Processing for Trigger Word Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Speech recognition systems face computational limitations in devices that monitor acoustic environments, making it challenging to detect trigger words efficiently, especially in devices with limited processing capacity.
Innovation Solution
Implementing a neural network-based waveform processing approach that uses a parameterized non-linear transformation to detect trigger words, involving a feature extractor and feature analyzer, and a Hidden Markov Model recognizer to identify the presence and timing of trigger words within sampled waveforms, with optional further classification for accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If speech recognition is performed using traditional methods, then recognition accuracy can be achieved, but computational cost becomes too high for devices with limited processing capacity
Solution Approach 1:
The speech recognition system is divided into two segments: a lightweight trigger word detection module that runs locally on the device with limited computational resources, and a more complex speech recognition module that runs on a remote server with sufficient computational power. This segmentation allows the device to perform basic trigger detection locally while offloading intensive recognition tasks to the server, thus resolving the contradiction between recognition accuracy and computational cost.
Solution Approach 2:
A remote server acts as an intermediary between the audio device and the speech recognition process. The device sends audio data to the server, which then performs the computationally intensive speech recognition and returns results to the device. This intermediary approach enables the device to achieve high recognition accuracy without bearing the full computational burden, effectively resolving the contradiction between accuracy and power consumption.
2Measurement precision
If all speech processing is performed on a remote server, then recognition accuracy improves, but latency increases due to network communication
Solution Approach 1:
The system performs preliminary trigger word detection locally on the device before sending audio data to the remote server. This preliminary action filters out non-trigger audio segments, so the device only transmits relevant audio data containing potential trigger words. This reduces the amount of data that needs to be communicated over the network and processed by the server, thereby reducing latency while maintaining recognition accuracy.
Solution Approach 2:
The speech processing task is segmented into local trigger detection and remote speech recognition. The local segment performs quick trigger word identification using a lightweight model, while the remote segment performs comprehensive speech recognition only when triggered. This segmentation reduces the time spent on network communication and server processing by avoiding transmission of all audio data, thus reducing latency while preserving accuracy.
3Power
If a lightweight model is used for local processing, then computational requirements are reduced, but detection accuracy decreases
Solution Approach 1:
The detection system is segmented into two levels: a lightweight local detector that identifies potential trigger words using minimal computational resources, and a more accurate remote recognizer that confirms and processes the speech. The local detector uses simplified models to reduce computational requirements, while the remote server handles the accuracy-critical recognition tasks, thus resolving the contradiction between computational capacity and detection accuracy.
Solution Approach 2:
The remote server acts as an intermediary that compensates for the limitations of the lightweight local model. The local device uses its computational constraints to perform initial filtering, and the remote server, with unlimited computational resources, performs the accurate recognition. This intermediary approach allows the system to use lightweight local processing without sacrificing overall detection accuracy, as the server handles the accuracy-critical tasks.
Data Source
AI summary
An approach to speech recognition, and in particular trigger word detection, implements fixed feature extraction form waveform samples with a neural network (NN). For example, rather than computing Log Frequency Band Energies (LFBEs), a convolutional neural network is used. In some implementations, this NN waveform processing is combined with a trained secondary classification that makes use of phonetic segmentation of a possible trigger word occurrence.