基于统计阈值的鲁棒性语音识别

近几十年来,语音识别系统已由实验室环境走向真实的世界中.在不同的环境噪声下,识别性能却仍不尽人意,尤其是在低信噪比的环境中.为解决在低信噪比情况下的低识别率的问题,以声学参数MFCC( Mel-frequen-cy cepstrum coefficient)为基础,提出了一种基于统计阈值的倒谱均值方差归一化算法,该算法...

Statistical thresholding for robust ASR

Speech recognition systems have been applied in real world applications for several decades, where there should be an unsatisfactory recognition performance under various noise conditions, particularly in lower signal-to-noise ratio (SNR) circumstances. In this paper, we propose a statistical thresholding method for mean and variance normalization technique, further reducing the mismatch between training and testing environments, which makes an automatic speech recognition system more robust to environmental changes. Mel-frequency cepstrum coefficient(MFCC) features are extracted as acoustic features, and they are further normalized with the mean and variance normalization method to get the cepstral mean and variance normalization (CMVN) features. The proposed statistical thresholding method is then applied. The viability of the proposed approach was verified in various experiments with different types of background noises at different SNR levels. In an isolated word recognition task, the experimental results show that the proposed approach reduced the error rate by over 40% in some cases compared with the baseline MFCC front-end, and under lower SNR conditions the proposed method also outperforms other robust features such as cepstral mean subtraction (CMS) and CMVN.