Abstract:

At present, many portable devices of information collection were designed and applied to agricultural product market information acquisition. Speech recognition has made great progress, and voice interface can be applied to these devices, but the general robust recognition algorithms are too complicated. In order to solve this problem, this paper proposed a robust speech recognition method, suitable for the devices operation environment. At first, MMSE amplitude estimator was used to enhance the noisy speech signal and improve signal-to-noise ratio (SNR) of the input signal; and then, for the speech distortion and residual noise after enhancement, the cepstrum mean and variance normalization (CMVN) methods were used to compensate. The experimental results showed that the combined algorithm could effectively improve the system correct recognition rate, especially under low SNR (0~10 dB) environment.

Key words: agricultural market information, speech recognition, minimum mean square error, feature compensation

摘要：

当前便携式农产品市场信息采集设备缺少语音接口，且通用领域的识别算法又过于复杂，为此提出一种适用于该设备作业环境的语音识别鲁棒性方法。首先利用MMSE估计器对带噪信号进行增强，以提高输入信号的信噪比；对增强后产生的语音失真和残留噪声，再利用倒谱均值方差归一化（CMVN）方法进行补偿。实验结果表明，该联合后的算法能有效的提高系统的识别率，特别是在低信噪比（0~10 dB）环境下更为有效。

关键词: 农产品市场信息；语音识别；最小均方误差；特征补偿