Hardware optimization and exploration of feature extraction and feature scoring for speech recognition
Hardware optimization and exploration of feature extraction and feature scoring for speech recognition(2009)
摘要
Speech is a natural form of communication, and speech recognition systems have been deployed in a wide range of applications and environments. However, speech recognition is extremely resource-intensive and future applications will only require more onerous performance and energy specifications without sacrificing recognition accuracy. Although this can be partially alleviated with better algorithms or faster processors, such approaches will not be able to deliver the orders of magnitude in improvement necessary. This work addresses this problem by implementing speech recognition in custom hardware, and introduces a hardware optimization methodology for the feature extraction and feature scoring stages of speech recognition. Moving from software to hardware improves decoding speed and energy consumption without sacrificing accuracy, but also increases the parameters to set and metrics to optimize for. First we profile a best quality speech recognizer to understand how a processor limits performance and the potential of custom hardware. Next we create a set of architecture exploration tools, describe how each design parameter affects the results, and generate a set of guidelines to make hardware optimization more efficient. We also convert software-based techniques intended to accelerate speech recognition into novel hardware to combine the performance gains from software and hardware. To demonstrate our optimization methodology and guidelines, we implement a hardware feature extraction and feature scoring virtual ASIC capable of recognizing speech at 28 times faster than real-time. This design is almost 10 times faster than software recognizers while using a 12 times slower clock.
更多查看译文
关键词
speech recognition,feature extraction,recognition accuracy,custom hardware,speech recognition system,hardware optimization methodology,best quality speech recognizer,hardware optimization,novel hardware,hardware feature extraction
AI 理解论文
溯源树
样例
生成溯源树,研究论文发展脉络