【摘要】 基于神经网络的语音增强算法相比于传统方法具有更好的语音增强效果,但因网络规模大导致其难以实时实施于助听器中。对此,本文提出了一种具有低复杂度的循环神经网络用于增强双耳语音。该算法结合双耳语音提取梅尔频率倒谱系数和耳间相位差作为网络的输入特征,在使用双耳语音振幅信息的基础上结合语音空间线索更好的映射目标语音和带噪语音之间的关系。实验结果表明,与助听器常用算法相比,本文算法的双耳平均信噪比和短时语音可懂度平均提高了4.68 dB和4.5%;与基于神经网络的算法相比,本文算法的双耳平均信噪比和短时语音可懂度分别提高了1.63 dB和4.8%。当时钟频率为10 MHz时,本文网络的硬件设计需要4.2 ms左右的处理时间,可以满足助听器的实时需求。更多还原

【Abstract】 Neural network-based speech enhancement algorithms provide better speech enhancement than traditional methods,but the large size of the network makes them difficult to implement in hearing aids in real time. In this paper,a recurrent neural network with low complexity is proposed for enhancing binaural speech. The algorithm combines binaural speech to extract Mel-frequency cepstral coefficient and interaural phase difference as input features to the network,and uses binaural speech amplitude information in combination with speech space cues to better map the relationship between target speech and noisy speech. The experimental results show that the average binaural signal-to-noise ratio and short-term speech intelligibility of this algorithm improve by an average of 4. 68 d B and4.5% compared to commonly used algorithms for hearing aids; compared to neural network-based algorithms,the average binaural signal-to-noise ratio and short-term speech intelligibility of this algorithm improve by 1. 63 d B and4.8% respectively. When the clock frequency is 10 MHz,the hardware design of this paper requires a processing time of about 4.2 ms,which can meet the real-time requirements of hearing aids.更多还原