End-point detecting method applied to speech identification system

Abstract

The invention relates to a terminal detection method for a voice recognition system. The method combines energy double-threshold algorithm with judgment by a model classifier, namely, a detection-verification strategy, to detect terminals of voice signals. In the detection process, energy is initially judged through energy double-threshold algorithm; in the verification process, firstly, a segmental zero-passage rate is judged and then the model classifier is adopted for further verification; the detection-verification strategy includes the following steps: model training, initial energy judging, zero-passage rate judging, model classifier judging, voice terminal determination and voice segment confirmation. The method can effectively and accurately locate the terminals of the voice; meanwhile, the method is easy to realize on line and is applicable to voice real time recognition systems.

Description

technical field [0001] The invention relates to the field of speech recognition. Background technique [0002] In the speech recognition system, the digital speech signal is composed of a mixture of silent segments, noise segments and speech segments. In this signal, speech is distinguished from various non-speech signals, and the accurate determination of the speech signal is called an endpoint. Detection or Voice Activity Detection (Voice Activity Detection, VAD). Whether the endpoint detection is correct or not will directly affect the performance of the speech recognition system, which is manifested in two aspects: accuracy and speed: first, if the silence and noise segments in the signal are removed, it will help the system to accurately extract speech features and improve speech recognition. Second, if the signal to be processed contains a large number of non-speech parts, an effective endpoint detection algorithm can remove the calculation of the noise segment, the c...

AI Technical Summary

Problems solved by technology

The method based on characteristics such as energy and zero-crossing rate has simple logic, high computational efficiency, and is easy to apply to real-time systems. This type of method works well in the case of high SNR, but the performance drops sharply in the case of low SNR. decline

The method based on the model classifier can achieve better results when the model matches the actual environment, but the disadvantage of this method is that there are many kinds of noisy environments, and it is impossible to establish corresponding models for all environments. In addition, based on The model method generally uses features with more dimensions, so the calculation efficiency is relatively low

Images