Speech Recognition Channel Verification for Noise Rejection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Automatic speech recognition systems perform poorly in noisy environments, particularly when background speakers utter in-vocabulary words, leading to errors such as insertions, which existing techniques like training non-speech and garbage models fail to effectively address.
Innovation Solution
The method involves determining scores for an input signal associated with speech recognition model states, computing a difference value between the input signal and an average signal, and processing the input signal based on this difference to differentiate between speech and noise channels, adjusting scores to bias probabilities, and using channel verification to reject noise and maintain accurate speech recognition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional speech recognition models are used in noisy environments, then speech recognition can be performed, but recognition accuracy deteriorates due to noise-induced insertions and errors
Solution Approach 1:
The patent segments the speech recognition task into two distinct stages: channel verification (determining whether the input signal passed through the speech channel) and speech recognition (performing actual recognition). This segmentation allows the system to first filter out noise-based signals before attempting recognition, thereby improving accuracy in noisy environments
Solution Approach 2:
The patent introduces an intermediary channel verification mechanism that acts as a gatekeeper between the noisy input signal and the speech recognition model. This intermediary stage computes channel deviation scores to determine whether the input likely passed through the speech channel, effectively filtering out noise-induced insertions before they reach the recognition model
2Object-affected harmful factors
If non-speech and garbage models are trained to reject environmental sounds, then some noise rejection is achieved, but they fail to reject in-vocabulary speech from background speakers
Solution Approach 1:
The patent changes the parameter being verified from speech content to channel characteristics. Instead of trying to recognize or reject specific sounds, the system verifies whether the signal passed through the speech channel by analyzing channel deviation in the cepstral domain, which inherently handles both environmental sounds and background speech uniformly
3Object-affected harmful factors
If channel verification is performed using cepstral analysis, then noise rejection improves, but computational complexity increases
Solution Approach 1:
The patent extracts only the essential channel verification computation from the full speech recognition pipeline. By focusing solely on computing channel deviation scores using cepstral analysis and comparing against a threshold, the system achieves effective noise rejection with minimal additional computational overhead, avoiding the need for complex full-model verification
Data Source
AI summary
A method for automatic speech recognition includes determining for an input signal a plurality scores representative of certainties that the input signal is associated with corresponding states of a speech recognition model, using the speech recognition model and the determined scores to compute an average signal, computing a difference value representative of a difference between the input signal and the average signal, and processing the input signal in accordance with the difference value.