Neural Network Beam Selection for Speech Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Speech recognition systems face challenges in identifying desired speech in noisy environments, where multiple speakers and background noise interfere with the ability to accurately process user commands, leading to difficulties in isolating the audio corresponding to the intended speaker.
Innovation Solution
A neural network classifier is trained to perform voice activity detection and determine the direction of desired speech, allowing the system to select and process audio from the appropriate direction, thereby improving speech recognition by isolating the desired speech and reducing noise interference.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional speech recognition processes all audio input, then speech recognition can be performed, but noise interference from multiple speakers and background sounds degrades recognition accuracy
Solution Approach 1:
The patent segments the audio input into multiple beams, where each beam captures audio from a specific spatial direction. The neural network then selects the beam containing the desired speech, effectively segmenting the audio processing to isolate the target speaker from noise and other speakers.
Solution Approach 2:
The patent applies local quality by training the neural network to identify and select specific beams with desirable audio characteristics. Each beam is evaluated locally for voice activity and speech quality, allowing the system to focus processing on the most relevant audio source while filtering out noise from other directions.
2Object-affected harmful factors
If beamforming is used to isolate speech from specific directions, then noise interference is reduced, but the complexity of determining which beam contains desired speech increases
Solution Approach 1:
The patent replaces complex mechanical or rule-based beam selection mechanisms with a neural network classifier. The neural network learns optimal beam selection through training on labeled data, substituting intricate decision logic with a trained model that automatically identifies the beam containing desired speech based on learned patterns.
Solution Approach 2:
The neural network performs self-service by automatically selecting the appropriate beam without requiring manual configuration or complex external control systems. The model independently evaluates all beams and selects the optimal one based on its trained understanding of speech characteristics, reducing the need for additional control complexity.
3Reliability
If all audio beams are processed for speech recognition, then no speech is missed, but computational resources are wasted on beams containing noise or non-speech audio
Solution Approach 1:
The patent applies preliminary action by using the neural network to pre-screen and select the most promising beam before performing full speech recognition processing. This preliminary selection step filters out beams containing only noise or non-speech audio, ensuring that computationally intensive speech recognition is applied only to beams with high probability of containing desired speech.
Solution Approach 2:
The patent uses partial action by processing only the selected beam containing desired speech rather than all audio beams. The neural network identifies and focuses computational resources on the single most relevant beam, performing partial processing that is sufficient for accurate speech recognition while avoiding the excessive computation that would result from processing all beams equally.
Data Source
AI summary
A neural network model, such as a deep neural network (DNN), is trained using many speech examples to perform beam selection in a microphone array-based speech processing system. The DNN is trained using many different speech examples that are labeled with position or direction information relative to a training microphone array. The DNN may then be trained to recognize a direction of incoming speech so that at runtime the trained DNN may process input audio data from a microphone array and may output to a beam selector an indicator of the desired beam that may be selected for further processing. The DNN may be configured to output a beam index and/or coordinates (or other position data) corresponding to an estimated location of the detected speech. The DNN may also be configured to output acoustic unit data corresponding to speech units (for example corresponding to phonemes, senons, etc. such as those of a detected wakeword or other word).