燕山大学学报
燕山大學學報
연산대학학보
JOURNAL OF YANSHAN UNIVERSITY
2015年
2期
133-138
,共6页
李荣芸%赵晓群%徐静云
李榮蕓%趙曉群%徐靜雲
리영예%조효군%서정운
模式识别%清/浊/静音判决%自适应阈值%低信噪比%低速率语音编码
模式識彆%清/濁/靜音判決%自適應閾值%低信譟比%低速率語音編碼
모식식별%청/탁/정음판결%자괄응역치%저신조비%저속솔어음편마
pattern recognition%unvoiced/voiced/silence detection adaptive threshold%low SNR%low bit-rate speech coding
清/浊/静音判决(UVS)是语音压缩、合成以及识别中的一个重要参数。为了解决传统判决方法训练过程复杂,导致语音编码效率低的问题,给出一种无训练过程的判决方法。提取基于循环平均幅度差的特征参量,利用判决参数间的相关性,自适应调整阈值,实现清/浊/静音判决。该判决方法具有很好的抗噪声干扰能力,有效提高判决的准确率。测试结果表明:该算法简化了清/浊/静音判决的计算量,清音误判率降低了10%,浊音误判率保持在4%以内;将该算法应用于低速率语音编码方案 MELP(mixed excitation linear prediction)0.6 kbps的清浊音判决中,解码后的合成语音质量优于原始 MELP 编码方案,PESQ 分数提高0.3,具有较好的可懂度和自然度。
清/濁/靜音判決(UVS)是語音壓縮、閤成以及識彆中的一箇重要參數。為瞭解決傳統判決方法訓練過程複雜,導緻語音編碼效率低的問題,給齣一種無訓練過程的判決方法。提取基于循環平均幅度差的特徵參量,利用判決參數間的相關性,自適應調整閾值,實現清/濁/靜音判決。該判決方法具有很好的抗譟聲榦擾能力,有效提高判決的準確率。測試結果錶明:該算法簡化瞭清/濁/靜音判決的計算量,清音誤判率降低瞭10%,濁音誤判率保持在4%以內;將該算法應用于低速率語音編碼方案 MELP(mixed excitation linear prediction)0.6 kbps的清濁音判決中,解碼後的閤成語音質量優于原始 MELP 編碼方案,PESQ 分數提高0.3,具有較好的可懂度和自然度。
청/탁/정음판결(UVS)시어음압축、합성이급식별중적일개중요삼수。위료해결전통판결방법훈련과정복잡,도치어음편마효솔저적문제,급출일충무훈련과정적판결방법。제취기우순배평균폭도차적특정삼량,이용판결삼수간적상관성,자괄응조정역치,실현청/탁/정음판결。해판결방법구유흔호적항조성간우능력,유효제고판결적준학솔。측시결과표명:해산법간화료청/탁/정음판결적계산량,청음오판솔강저료10%,탁음오판솔보지재4%이내;장해산법응용우저속솔어음편마방안 MELP(mixed excitation linear prediction)0.6 kbps적청탁음판결중,해마후적합성어음질량우우원시 MELP 편마방안,PESQ 분수제고0.3,구유교호적가동도화자연도。
The Unvoiced/ Voiced/ Silence detection UVS provides a preliminary acoustic segment which is a key parameter in speech compression synthesis and recognition.The complication of traditional UVS methods?? training procedure causes low efficiency of speech vocoder.To solve this problem a UVS detection without training proceeding is proposed in this paper.After new characteristic parameters of unvoiced and voiced signal are extracted adaptable threshold is proposed based on the correlation of those parameters.With its perfect an?ti?noise ability the correct rate of this detection improves sharply.The simulation result shows that this algorithm not only simplifies the unvoiced/ voiced/ silence detection but also efficiently decreases 10% of unvoiced and maintains lower than 4% of voiced discrimination error.The improved 0.6 kbps MELP vocoder applying this detection algorithm gets a 0.3 higher PESQ score and better synthetic speech performance compared with original vocoder which produces good natural and intelligible speech.