Selective Noise Suppression for Automatic Speech Recognition

Solution Overview

Problem

Existing automatic speech recognition systems in vehicle environments face challenges in maintaining accuracy due to inappropriate noise suppression, which can introduce musical distortions and reduce system performance, especially when the signal-to-noise ratio (SNR) of speech frames is high.

Innovation Solution

A method is introduced where audio signals are processed by identifying voiced-frames with a high signal-to-noise ratio (SNR) and bypassing noise suppression for those frames, while applying noise suppression to frames with lower SNR, thereby improving the accuracy of automatic speech recognition without introducing undesirable distortions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Object-affected harmful factors

If noise suppression is applied to all audio frames, then noise reduction is achieved, but musical distortions are introduced and ASR accuracy deteriorates for high SNR frames

Engineering Contradiction:

ImprovenoiseVSAvoidASR accuracy

Core Design Contradiction:

Object-affected harmful factorsVSManufacturing precision

Solution Approach 1:

The patent applies different noise suppression treatments to different portions of the audio signal based on local characteristics. Specifically, voiced frames with high SNR (greater than threshold TV1) are excluded from noise suppression to preserve speech quality, while other frames (unvoiced frames and voiced frames with low SNR) continue to receive noise suppression. This local differentiation resolves the contradiction by applying noise suppression only where it is beneficial.

Inventive Principle:

Principle #3Local quality

Solution Approach 2:

The patent changes the parameter of noise suppression application from a constant state (applied to all frames) to a variable state (selectively applied based on frame type and SNR). By introducing conditional logic that evaluates SNR values and frame classification, the system dynamically adjusts whether noise suppression is applied, thereby optimizing both noise reduction and ASR accuracy across different signal conditions.

Inventive Principle:

Principle #35Parameter changes

2Manufacturing precision

If noise suppression is bypassed for high SNR voiced-frames, then musical distortions are avoided, but noise may remain in low SNR frames

Engineering Contradiction:

Improvespeech qualityVSAvoidnoise

Core Design Contradiction:

Manufacturing precisionVSObject-affected harmful factors

Solution Approach 1:

The patent applies different noise suppression treatments to different portions of the audio signal based on local characteristics. Specifically, voiced frames with high SNR (greater than threshold TV1) are excluded from noise suppression to preserve speech quality, while other frames (unvoiced frames and voiced frames with low SNR) continue to receive noise suppression. This local differentiation resolves the contradiction by applying noise suppression only where it is beneficial.

Inventive Principle:

Principle #3Local quality

Solution Approach 2:

The patent changes the parameter of noise suppression application from a constant state (applied to all frames) to a variable state (selectively applied based on frame type and SNR). By introducing conditional logic that evaluates SNR values and frame classification, the system dynamically adjusts whether noise suppression is applied, thereby optimizing both noise reduction and ASR accuracy across different signal conditions.

Inventive Principle:

Principle #35Parameter changes

3Manufacturing precision

If selective noise suppression is implemented based on SNR thresholds, then ASR accuracy is improved, but processing complexity increases

Engineering Contradiction:

ImproveASR accuracyVSAvoidprocessing complexity

Core Design Contradiction:

Manufacturing precisionVSDevice complexity

Solution Approach 1:

The patent segments the audio signal processing into distinct categories: voiced frames, unvoiced frames, and silence frames. Each segment is evaluated independently using SNR thresholds (TV1 for voiced, TU1 for unvoiced) to determine whether noise suppression should be applied. This segmentation approach manages complexity by organizing the selective processing logic into clear, manageable categories rather than treating all frames uniformly.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent changes the parameter of noise suppression application from a constant state (applied to all frames) to a variable state (selectively applied based on frame type and SNR). By introducing conditional logic that evaluates SNR values and frame classification, the system dynamically adjusts whether noise suppression is applied, thereby optimizing both noise reduction and ASR accuracy across different signal conditions.

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentUS9830925B2Selective noise suppression during automatic speech recognition

Publication Date: 2017.11.28 GM GLOBAL TECHNOLOGY OPERATIONS LLC

AI summary

An automatic speech recognition engine and a method of using the engine is described. The method pertains to front-end processing an audio signal and includes the steps of: identifying a plurality of voiced-frames of the audio signal; determining that one or more of the plurality of voiced-frames have a signal-to-noise (SNR) value greater than a first predetermined threshold; and based on the determination, bypassing noise suppression for the one or more of the plurality of voiced-frames.