Hardware Implementation Based On Low-resource Speech Recognition System

As an important branch of artificial intelligence machine learning,language recognition technology has an important position in Internet of Things technology and software development,and in ordinary acoustic models,Upon a low-resource database condition,traditional acoustic GMM-HMM model can't achieve a satisfying recognition rate and has large parameter scale.In order to solve those problems,a speech recognition BN-SGMM-HMM model is proposed in this article.In the acoustic feature aspect,a DNN-based BN(Bottle Neck)feature is extracted which improves the system's discriminability and robustness capability;meanwhile,the Dropout strategy is employed to prevent over-fitting problem during the training process.In the acoustic model aspect,the SGMM(Subspace Gauss Mixture Model)is adopted to decrease the parameter scale.The improvements in these two aspects have also improved the recognition rate of low-resource speech recognition systems.The experiments in this paper prove that the implement of BN-SGMM-HMM low-resource speech recognition system can train the better recognition effect under limited training corpus.While In the hardware implementation part,the open-source Chinese corpus is used for training based on the BN-SGMM-HMM acoustic model,and the trained acoustic model is implemented on the Raspberry Pi,and the microphone is used as the voice input through the Kaldi internal decoder.Recognize the input voice,and finally display the recognition result on the terminal.The innovation of the language recognition system lies in:In terms of software development,the BN-SGMM-HMM acoustic model is used as the basic model and the Kaldi speech recognition toolkit is used to train the model,and has internal feature extraction scripts and language model generation tools,which has changed the need for experienced engineers in speech recognition development in the past.This situation reduces the cycle of speech recognition system developers;in terms of hardware migration,because the hardware implementation uses the open source hardware Raspberry Pi,the user is extensive and the internal environment is open source,compared to other ARM-based development boards and ASICs Reduce the cost of development cycle and tape out.