HMM非特定人連續(xù)語(yǔ)音識(shí)別的嵌入式實(shí)現(xiàn)
Embedded Implementation of HMM Speaker-Independent Continuous Speech Recognition System
-
摘要: 嵌入式系統(tǒng)正逐漸成為語(yǔ)音識(shí)別實(shí)際應(yīng)用的首選平臺(tái)。該文在嵌入式平臺(tái)上研究HMM連續(xù)語(yǔ)音識(shí)別的計(jì)算復(fù)雜度要素,提出特征系數(shù)屏蔽方法和綜合剪枝相結(jié)合的瘦身計(jì)算方法,降低計(jì)算復(fù)雜度并保持識(shí)別率。該方法在嵌入式平臺(tái)上研究的實(shí)驗(yàn)數(shù)據(jù)表明,HMM連續(xù)語(yǔ)音識(shí)別瘦身系統(tǒng)與基線系統(tǒng)相比,計(jì)算時(shí)間從基線系統(tǒng)的100％降低到27.91％,識(shí)別率僅從基線系統(tǒng)的89.65％下降到89.41％。
-
關(guān)鍵詞:
- 嵌入式系統(tǒng); 語(yǔ)音識(shí)別; 搜索算法; 特征屏蔽
Abstract: The embedded systems are gradually becoming the first choice of platforms which should be used for real-time speech recognition system. This paper discusses the computation complexity factors of HMM-based continuous speech recognition for embedded system. An optimized way integrating feature masking and pruning is presented to reduce the computation complexity and keep the recognition accuracy. The experiments for embedded system show that, comparing with the base-line system, the computation time is reduced from 100% to 27.91%, and the recognition accuracy is degraded only from 89.65% to 89.41%. -
Du Limin.[J].Feng Junlan, Song Yi, Sun Jinchen. A ChineseEnglish speech translation prototype system: CEST-CAS1.0.ICSPAT99, Orlando, USA.1999,:-[2]Du Limin, Feng Junlar, Song Yi, Wang Heng. Speech translation on internet CEST-CAS2.0. Proc. of ISIMP2001, Hong Kong,2001: 189- 192.[3]Rabiner L, Juang B H. Fundamentals of Speech Recognition.New Jersey, USA, Prentice Hall, 1993:350 - 352.[4]Ney H, Ortmanns S. Dynamic programming search for continuous speech recognition[J].IEEE Signal Processing Magazine.1999, 16(5):64- -
計(jì)量
- 文章訪問(wèn)數(shù): 3693
- HTML全文瀏覽量: 294
- PDF下載量: 69960
- 被引次數(shù): 0
下載:
