Speech recognition system and method using noise padding and normalization in dynamic time warping

Abstract

Speech recognition uses a wide token builder (66), gain and noise adapter (70) and noise adapted Dynamic Time Warping (60). Wide token builder produces a padded test token expanded with at least one blank frame before and after the input test utterance. Gain and noise adapter adapts each padded reference template with noise and gain qualities producing adapted reference templates having noise frames wherever a blank frame was originally placed and noise adapted speech where speech exists. Dynamic Time Warping (DTW) is performed on the noise adapted templates.

Description

field of invention [0001] The present invention relates generally to speech recognition, and more particularly to specific speaker recognition techniques in noisy environments. Background of the invention [0002] Speech recognition in noisy environments is a long-studied and still unsolved task. This task is characterized by the following parameters: [0003] 1. The recognition is different from person to person, and the reference template is generated by the speaking tone of the user in a specified "training dialogue"; [0004] 2. It is desirable to reduce the number of training tones to a minimum number (1-3), which, in the prior art, would make the Dynamic Time Warping (DTW) matching algorithm more efficient than the Hidden Markov Model (HMM) algorithm: [0005] 3. The phrase to be recognized is a single word; [0006] 4. The training stage is relatively low-noise, but additional environmental noise needs to be dealt with in the recognition; [0007] 5. The ambient n...

AI Technical Summary

Problems solved by technology

[0005] 3. The phrase to be recognized is a single word;

[0007] 5. The ambient noise is unknown to the system until the instant the user presses the talk (PTT) button to start speaking;

[0009] 7. The system has only limited fast access memory, it is impossible to run the DTW matching algorithm against all reference templates in a real-time and word recognition manner

[0012] 2. In the recognition stage, the word endpoint estimated by VAD is not accurate

[0028] The above approach has two disadvantages: (a) it cannot be adopted for the tasks defined in items (2) to (7) above; and (b) the one-shot approach suffers from an efficiency loss in dealing with acoustic mismatch ( Problem 1), since there is no precise information on the noise level in the primary algorithm

This creates a problem with the exclusion mechanism, because double-syllable intonation is more likely to be excluded than single-syllable intonation

Images