论文部分内容阅读
在嵌入平台上实现高性能的汉语数码语音识别(MDSR),对于电话通讯、工业控制等都具有极高的实用价值。该文描述了一个在16bit定点DSP芯片上实现的高性能汉语数码语音识别系统。识别模型采用连续隐Markov模型(CHMM),识别特征采用Mel频标倒谱系数(MFCC)。在模型的训练中引入MCE区分性训练进一步提高了系统的识别性能。识别过程采用单级识别框架,降低了芯片上系统部分的复杂性,同时保证了很高的识别性能与稳健性。实验证明该系统对11汉语数码发音可以达到98.3%的识别正确率,在58.5MIPS的16bit定点DSP上进行一次识别只需要35ms。
The realization of high performance Chinese digital speech recognition (MDSR) on the embedded platform has extremely high practical value for telephone communication and industrial control. This article describes a high-performance Chinese digital speech recognition system implemented on a 16-bit fixed-point DSP chip. The identification model adopts the continuous hidden Markov model (CHMM) and the recognition feature adopts the Mel Cc cepstrum coefficient (MFCC). The introduction of MCE discriminative training in model training further improves the recognition performance of the system. The identification process uses a single-stage identification framework, reducing the complexity of the system components on the chip, while ensuring high recognition performance and robustness. Experiments show that the system can achieve a recognition rate of 98.3% on 11 Chinese digital utterances. Only 35ms is needed to recognize a 58bit MIPS on a 16bit fixed-point DSP.