Per-Channel Energy Normalization for Far-Field Speech Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current voice-activated electronic devices face challenges in robust speech recognition, particularly in far-field and noisy environments due to background noise and varying loudness levels, which affects the reliability of hands-free communication.
Innovation Solution
The implementation of a per-channel energy normalization (PCEN) frontend, which includes automatic gain control-based dynamic compression, is used to distinguish targeted audio from background noise and stabilize signal levels, enhancing recognition performance without increasing model complexity or inference-time cost.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Length of moving object
If traditional speech recognition systems are used in far-field conditions, then the device can capture voice inputs from a distance, but the speech recognition accuracy deteriorates due to signal attenuation and background noise
Solution Approach 1:
The patent applies per-channel energy normalization (PCEN) which dynamically adjusts the energy parameters of different frequency channels. This transforms the audio signal by normalizing energy across channels, making the speech signal more robust to attenuation and noise while preserving the relative spectral characteristics needed for accurate recognition
Solution Approach 2:
The system performs preliminary noise mitigation and energy normalization on the audio signal before it reaches the speech recognition model. By pre-processing the signal to compensate for expected attenuation and noise effects, the system prepares the input data in advance to maintain recognition accuracy in far-field conditions
2Reliability
If noise mitigation measures are implemented to improve speech recognition in noisy environments, then the recognition robustness improves, but the device complexity increases
Solution Approach 1:
The patent modifies the energy parameters of frequency channels through per-channel normalization, changing the statistical properties of the signal to enhance noise robustness. This parameter transformation approach provides effective noise mitigation through relatively simple computational operations on the spectral features
Solution Approach 2:
The system introduces an intermediate processing stage (PCEN) between audio capture and speech recognition that acts as a mediator. This intermediate layer transforms the raw audio into a normalized representation that is more robust to noise, while the computational complexity remains manageable through efficient implementation of the normalization algorithm
3Reliability
If per-channel energy normalization is applied to improve far-field speech recognition, then the recognition performance improves, but the computational processing time increases
Solution Approach 1:
The patent applies per-channel energy normalization which transforms the audio signal by adjusting energy parameters across frequency channels. This parameter transformation is computationally efficient as it operates on spectral features rather than raw waveforms, providing far-field robustness with minimal additional processing time
Solution Approach 2:
The system applies a simplified form of normalization that processes only the essential energy parameters needed for far-field robustness rather than performing exhaustive signal processing. This partial action approach achieves the necessary performance improvement without the full computational cost of more comprehensive processing schemes
Data Source
AI summary
The various implementations described herein include methods, devices, and systems for recognizing speech, such as user commands. In one aspect, a method includes: (1) receiving audio input data via the one or more microphones; (2) generating a plurality of energy channels for the audio input data; (3) generating a feature vector by performing a per-channel normalization to each channel of the plurality of energy channels; and (4) obtaining recognized speech from the audio input utilizing the feature vector.