Audio Pre-Conditioning for Automatic Speech Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing audio processing systems optimized for human perception often degrade the performance of machine perception systems due to differing requirements and parameters, leading to suboptimal performance in automatic speech recognition.
Innovation Solution
Implementing pre-conditioning techniques that allow for increased echo leakage, background noise, and masking noise to improve machine perception, using parameters such as faster echo cancellation and noise injection tailored for machine perception systems.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If audio processing parameters are optimized for human perception (e.g., aggressive echo cancellation with 100-200ms convergence, high noise suppression target more than 20 dB below speech level), then human listening experience is improved, but machine perception performance deteriorates
Solution Approach 1:
The patent segments the audio processing pipeline into two distinct processing paths: one optimized for human perception and another optimized for machine perception. This allows each path to use parameters tailored to its specific requirements without compromising the other, resolving the contradiction between human listening experience and machine perception performance
Solution Approach 2:
The patent applies different quality characteristics to different processing streams by introducing masking noise specifically for the machine perception path while maintaining clean processing for human listening. This local differentiation allows aggressive noise suppression for machines while preserving natural audio quality for humans
2Object-affected harmful factors
If echo cancellation convergence is set to 100-200ms for human perception, then echo is effectively removed for human listeners, but machine perception accuracy decreases
Solution Approach 1:
The patent applies preliminary noise addition to the audio signal before it reaches the machine perception processing stage. This preliminary action ensures that the machine perception system receives pre-conditioned audio with appropriate noise characteristics, improving recognition accuracy without affecting the echo cancellation performance for human listeners
3Object-affected harmful factors
If noise suppression target is set more than 20 dB below speech level for human perception, then background noise is significantly reduced for human listeners, but automatic speech recognition performance deteriorates
Solution Approach 1:
The patent inverts the conventional approach by adding noise rather than removing it for machine perception. This counterintuitive strategy improves automatic speech recognition by providing the machine with noise characteristics that match its training conditions, while the human perception path continues to benefit from aggressive noise suppression
Data Source
Figure 1~2
Figure 3
Figure 4~5
AI summary
An apparatus and method of pre-conditioning audio for machine perception. Machine perception differs from human perception, and different processing parameters are used for machine perception applications (e.g., speech to text processing) as compared to those used for human perception applications (e.g., voice communications). These different parameters may result in pre-conditioned audio that is worsened for human perception yet improved for machine perception.