End-to-End Speech Recognition Neural Network for Far-Field Audio
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional speech recognition systems face challenges in far-field environments due to low signal-to-noise ratios and inconsistent optimization objectives between front-end speech enhancement and back-end recognition, leading to poor accuracy and reliability.
Innovation Solution
An end-to-end neural network-based model for integrated speech enhancement and recognition that extracts enhanced features from multi-channel voice signals, optimizing both speech enhancement and recognition objectives jointly, using a complex convolutional neural network to perform beamforming and feature extraction while preserving phase information, and simulating far-field voice signals in real-time for enriched training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If conventional speech enhancement and recognition are used separately, then each module can be optimized independently, but the overall recognition accuracy in far-field environments deteriorates due to inconsistent optimization objectives
Solution Approach 1:
The patent merges the speech enhancement module and speech recognition module into a unified end-to-end neural network model. This integration allows the optimization objectives of both modules to be aligned, as they now share a common loss function and training process, thereby resolving the contradiction between independent optimization ease and overall recognition accuracy.
2Adaptability or versatility
If far-field speech recognition is performed in noisy environments, then the system can operate in real-world scenarios, but the signal-to-noise ratio decreases leading to poor recognition performance
Solution Approach 1:
The patent applies preliminary speech enhancement actions within the end-to-end model by incorporating convolutional neural network layers that perform beamforming and feature extraction before the main recognition process. This preliminary enhancement of speech signals in noisy environments improves the signal-to-noise ratio, thereby enhancing recognition reliability while maintaining adaptability to real-world scenarios.
3Measurement precision
If deep learning technologies are applied to speech recognition, then recognition accuracy improves, but the computational complexity and resource requirements increase
Solution Approach 1:
The patent segments the speech recognition system into distinct functional layers within the neural network, including convolutional layers for feature extraction, beamforming operations for spatial filtering, and recognition layers for text output. This segmentation allows for optimized computation at each stage, reducing overall computational complexity while maintaining high recognition accuracy through specialized processing in each segment.
Data Source
AI summary
The disclosure provides a speech recognition method, a device and a computer-readable storage medium. The method includes obtaining a first voice signal collected from a first microphone in a microphone array and a second voice signal collected from a second microphone in the microphone array, the microphone array including at least two microphones, such as two, three or six microphones. The method further includes extracting enhanced features associated with the first voice signal and the second voice signal through a neural network, and obtaining a speech recognition result based on the enhanced features extracted.