基于Kaldi的AI语音识别在嵌入式系统中的应用研究

Research on AI Speech Recognition Based on Kaldi in Embedded System

Kaldi为目前主流桌面端语音识别的人工智能框架,随着智能家居产品语音识别的需求增长,针对嵌入式硬件实现语音识别显得十分重要。针对Kaldi进行交叉编译并实现了基于ARM Cortex-A72内核的Raspberry Pi 4B嵌入式平台的移植,结合ReSpeaker 2-Mics Pi HAT,使用深度神经网络隐马科夫模型,实现了嵌入式实时离线大词汇量连续语音识别。实验结果表明,Kalid在嵌入式设备上运行语音识别算法时,并非预期的增加语音识别算法使用的CPU核心并行数有利于语音识别的响应时间。由于受制于算法框架和硬件资源的限制,应选择适合硬件条件的核心数来并行运算语音识别算法最佳,从而保证语音识别的速度。

This paper implements cross-compilation for Kaldi and the transplantation of the Raspberry Pi 4B embedded platform based on the ARM Cortex-A72 core,combined with ReSpeaker 2-Mics Pi HAT,using a deep neural network hidden Markov model,to achieve an embedded real-time offline large-scale Vocabulary continuous speech recognition.Experimental results show that when Kalid runs a speech recognition algorithm on an embedded device.It is not as expected that increasing the CPU core parallel number used in speech recognition algorithm is conducive to the response time of speech recognition.Due to the limitation of the algorithm framework and hardware resources,the core number parallel operation speech recognition algorithm suitable for hardware conditions should be selected to ensure the speed of speech recognition.