应用于语音识别片上系统的语音检测算法

A Speech Detection Algorithm for Speech Recognition Systems on Chip

语音识别技术的研究已经进入实用化阶段,而实用化语音识别系统中的一个关键技术就是可靠的语音检测。本文提出了一种基于有限状态机模型的实时语音检测算法(FSM-SD)。采用对数最大似然判决帧能量检测器和过零率检测器控制各状态之间的跳转关系。针对语音识别中的MFCC(Mel频标倒谱系数)和LPCC(线性预测倒谱参数)特征提取过程,分别得到两种不同的帧能量计算方法。将FSM-SD应用到在OAK DSP上实现的小词表汉语语音识别系统,通过实验验证了其对系统识别性能和噪声稳健性的有效保证。

With the development of speech recognition, a robust speech detector has been the integral part of the practical speech recognition system. In this paper we propose a new finite state machine (FSM) based speech detection algorithm. The inputs of the FSM are derived from a zero-crossing detector and a LML (Logarithm Maximum Likelihood) frame-energy detector, where two kinds of noise robust energy are respectively used for MFCC and LPCC. Based on the proposed scheme, a small-vocabulary mandarin speech recognition system on OAK DSP can give real-time accurate speech recognition result. Experiments have been conduct to verify the viability of the proposed algorithm.