基于自回归模型的加性噪声环境稳健语音识别

为提高噪声不平稳或不可估的情况下语音识别的稳健性,提出了利用自回归模型和短时平稳性假设,估计干净与噪声环境的语音数据,建立相应的语音识别模型,以达到抗噪效果的稳健语音信号处理方法。在N o iseX-92的4种噪声环境(w h ite,babb le,vo lvo,destroyer eng ine)从0到20 dB的不同信噪比下的“863”大词汇连续语音标准数据库的平均识别结果表明,该方法能够使得基于段长分布的隐M arkov模型的语音识别系统在25候选时声学层的音节相对错误率下降达到10.85%以下,同时相对正确识别率上升12.13%。

Autoregressive model-based robust speech recognition in additive noise environment

The robustness of speech recognition in a non-stationary or unknown additive noise environment is improved by a speech signal processing method based on an autoregressive(AR) model using the short time stationary assumption.The method captures the model-based speech information in both clean and noisy speech signals using a hidden Markov model(HMM) with the AR-estimated speech signal.Experimental results using the standard "863" large vocabulary continuous speech database with four noise(white,babble,volvo,destroyer engine) from NoiseX-92 at signal-to-noise ratios(SNR) from 0 to 20 dB show that the AR-model processing reduces the acoustic-level relative syllable error rate by 10.9% and improves the relative syllable recognition rate by 12.1% when using 25-candidate selection in the duration distribution-based hidden Markov model speech recognition system.