基于TMS320C6701EVM板的快速說(shuō)話人識(shí)別系統(tǒng)

本文在TMS320C6701EVM 板的基礎(chǔ)上實(shí)現(xiàn)一種快速的說(shuō)話人識(shí)別系統(tǒng)。本文提出一種基于段級(jí)語(yǔ)音特征的說(shuō)話人識(shí)別的快速算法，該算法在傳統(tǒng)的GMM 算法的基礎(chǔ)上使用段級(jí)語(yǔ)音特征對(duì)測(cè)試語(yǔ)音進(jìn)行數(shù)據(jù)量壓縮，以減少計(jì)算時(shí)間。并基于車比雪夫和不等式提出了基于協(xié)方差模型的段級(jí)特征的失真測(cè)度描述。本文根據(jù)實(shí)驗(yàn)選擇了段級(jí)特征語(yǔ)音段長(zhǎng)度，實(shí)驗(yàn)表明該算法在不顯著影響識(shí)別率的基礎(chǔ)上有效地減少了算法延遲，提高了識(shí)別速度。
關(guān)鍵詞:段級(jí)特征; 車比雪夫和不等式; 說(shuō)話人識(shí)別; 數(shù)字信號(hào)處理芯片
Abstract: This paper realize a fast speaker recognize system based on TMS320C6701 EVM board. This paper proposes a fast speaker recognize algorithm based on utterance level feature. Within the system, the algorithm’s framework is GMM and compress data using utterance level feature. The paper also gives out utterance level covariance distortion measure by using Chebyshev’s summation inequalities. This paper chooses the best utterance according to the experiments. Experiments results show that this algorithm efficiently reduce computation delay and fast recognition speed.
Key words: Utterance level feature; Chebyshev’s summation inequalities; Speaker recognition; DSP