Robust HMMs against environmental variation for speech recognition

(1)本课题研究的是如何使HISTORY算法对信噪比的变化具有鲁棒性。本文提出了一种用于HMM的噪声方差扩展技术，它包括简单地扩展HMM组成中噪声模型的倒谱系数的方差。使用NOISEX-92噪声数据，通过与说话人无关的数字识别测试，检查这种技术的效果。结果表明，0阶倒谱的方差扩展极大地提高了标准HMM对大范围信噪比失配的鲁棒性。适当的扩展因子被确定为与噪声类型无关，使得第零倒谱的扩展方差相对于其几何平均值约为5至6dB。(2)抗加性噪声的抗差谱分析研究本项目的一部分还发展了一种简单有效的时域方法来估计梅尔频率轴上的全轮询模型（Mel-LPC）。与传统的线性预测分析相比，这种方法只需要两倍的计算成本。性别相关的音素识别测试表明，梅尔LPC倒谱实现了一个显着的改善，在识别精度比传统的LP梅尔倒谱和梅尔频率倒谱系数（MFCC）。此外，噪声字识别测试表明，梅尔LPC倒谱是强大的宽带加性噪声比传统的LP梅尔倒谱和MFCC。

(1) A Study on Variance Expansion of HMMs Robust to Environmental VariationThis project addresses the problem of making HMMs robust to variation of SNR. This study developed a noise varicance expansion technique for HMMs, which consists of simply expanding the variace of cepstral coefficients for the noise model in HMM composition. The effect of this technique is examined through speaker independent digit recognition tests using NOISEX-92 noise data. The results show that the variance expansion of the 0th order cepstrum extremely improves robustness to a wide range of SNR mismatch over the standard HMM. The appropriate expansion factor is determined irrespective of noise types such that the expanded variance of the zeroth cepstrum is around 5 to 6dB with respect to its geometric mean.(2) A Stuty on a Robust Spectral Analysis to Additive NoiseA part of this project also developed a simple and efficient time domain technique to estimate an all-poll model on a mel-frequency axis (Mel-LPC). This method requires only two-fold computational cost as compared to conventional linear prediction analysis. Gender-dependent phoneme recognition tests show that the Mel-LPC cepstrum attains a significant improvement in recognition accuracy over conventional LP mel-cepstra and the mel-frequency cepstrum coefficients (MFCC). Furthermore, noisy word recognition tests revealed that the Mel-LPC cepstrum is robust to wide-band additive noise over conventional LP mel-cepstrum and MFCC.