573294 A7 B7 五、發明說明（） 發明領域 (請先閲讀背面之注意事項再填寫本頁) 本案係為一種向量量化的方法’尤指一 種非線性分頻向量量化的方法。 發明背景 在現今數位多媒體與網際網路快速成長 的時代，在有限的傳輸頻寬環境下，必須發 展更有效的語音編解碼方式，以便在傳輸頻 寬壅塞時，能提供良好的通信服務品質及使 用率，不致有斷訊的情形。而在人造衛星及 個人無線行動通訊應用上，如何以較低的位 元速率達到通訊品質的需求，也是值得研究 的課題。 經濟部智慧財產局員工消費合作社印製 為了因應未來二十一世紀多用途、多功 能的通訊時代，國際動態影像專家群組織 (Moving Picture Experts Group, MPEG)訂 定了 MPEG-4音頻協定，在語音編碼技術方 面，低位元率（4 . 8 k b p s以下）的”諧波及向 量激發信號編碼器”（H a r m ο n i c V e c t 〇 r Excitation Coding, HVXC)在 4 kbps 及 2 kbps 兩種固定位元率，所得到之合成語音品質均 高於現今位元率為4.8 kbps的FS1016編碼 方式，合成語音信號之品質達到”通訊品 質 ’’（Communication Quality)之需求；此外 HVXC編碼器亦採用可調式位元率編解碼架 構（位元率1.2k〜1.7kbps)，並在解碼端提 3 本紙張尺度適用中國國家標準（CNS)A4規格（210 X 297公釐） 573294 A7 B7 經濟部智慧財產局員工消費合作社印製 五、發明說明（） 供可變基週（Pitch)及可變撥放速度之功 能。由此可知，Η V X C語音編解碼系統在最近 將獲得重視與發展，並在通訊及數位化儲存 等方面，可能被廣泛的採用。 Η V X C語音編解碼器所採用的編碼方法 與”多頻帶激發信號語音編碼 法 ”（Multi-Band Excitation，ΜΒΕ)編碼方法 相似，此類言皆波編碼（H a r m ο n i c c o d i n g )方 式，對語音信號中有聲的（V o i ce d )音框，採 用諧波頻譜振幅向量量化之編碼方式；由於 語音諧波之假數與該音框語音之基週有關， 所以必須對具時變維度特性之頻譜封包信號 作向量量化，其編碼演算法不僅耗用大量之 計算，而且諧波編碼方式的合成語音品質決 定在於諧波頻譜封包向量量化品質之優劣， 所以如何簡單且有效的量化其封包向量，對 該編碼方式是極重要且有價值之研究主題。 而在實際運用時，降低計算量之結果，亦有 助於在個人電腦或DSP(Digital Signal P r o c e s s o r )微處理晶片上，以軟體或硬體將 編碼器與解碼器同時實現在一顆晶片上，執 行即時（r e a 1 - t i m e )編解語音的工作。 HVXC語音編碼器，有聲（Voiced)或無聲 (U n v 〇 i c e d )之語音採用不同之編碼方式；對 無聲語音使用傳統線性預估碼激式編碼架 構，而對於有聲語音採用類似多頻帶激發信 4 本紙張尺度適用中國國家標準（CNS)A4規格（210 X 297公釐） ---------- - 裝--------訂--------- (請先閱讀背面之注意事項再填寫本頁) 經濟部智慧財產局員工消費合作社印製 573294 . A7 五、發明說明（） 號語音編碼法之編碼方式。其編碼方式為對 輸入的每個音框信號，先求取其線性預估參 數，然後將該音框之語音通過線性預估分析 濾波器所得到之殘值信號，以2 5 6點 DFT(Discrete Fourier Transform)求得其殘值 信號之頻譜及基週值（Pi tch Lag)。對有聲音 才匡的語音信號，為傳送此殘值信號頻譜，必 需先對基頻（Fundamental Frequency)的每 一諧波頻帶，估測其諧波頻譜之振幅，再將 所有諧波頻譜振幅所組成之頻譜封包向量 (Spectral Envelope Vector，SEV)，以頻譜 封包向量量化（Spectral Envelope Vector Quantization，SEVQ)方式找出最佳的碼字及 增益並傳送最佳碼字索引值到解碼端。 由於頻譜封包向量之維度與該音框語音 之基週大小成正比，所以頻譜封包向量量化 所要量化之頻譜封包向量維度是時變的，（對 8 k Η z取樣之語音信號，其頻譜封包向量維度 Ρ範圍為10〜67)，為解決此一可變維度向量 量化（Variable-Dimension Vector Quantization，VDVQ)之問題，HVXC、語音編碼 器將頻譜封包向量以線性内插方式，轉換成 固定維度Μ ( = 4 4 )之頻譜封包形狀向量 f叶⑺= 1，2，···Μ}(步驟11)後，再以r( = 8)位元 向量碼薄作向量量化。 請參閱第一圖，其係習用諧波及向量語 5 本紙張尺度適用中國國家標準（CNS)A4規格（210 X 297公釐） ---------------I----訂 --------線 (請先閱讀背面之注意事項再填寫本頁) 573294 A7 B7 五、發明說明（ 音編解碼器之頻譜封包向量量化動作示意 圖。兩個各具有K( = 2；?/2= 1 6 )個之碼字向量之 碼薄（第一向量碼薄^^^Γ，^0，1，··.，^} Η及 第二向量碼薄15，向量量化 時，任一碼字均由此兩個碼薄各擇一個碼字 所組成，總共有L ( = 2 5 6 ) 種可能之碼字組 合；若以/.。及h分別代表碼薄P(F)及之索弓I 值，固定維度之碼字向量#，7·=尺5%+A，如方程 式（1 )所示： 若以％，/ = 0，1，...，10表示目前編碼音框之線性 預估係數，則聽覺加權濾波器（P e r c e p t u a 1 Weighting Filter)16 如下戶斤示： 10Σα，/Ζ_/ ! W{z) (請先閱讀背面之注意事項再填寫本頁) 經濟部智慧財產局員工消費合作社印製 10 10 \ u / Σ^2^'1' Σα^ /=0 /=0 其中，6與h之數值，在HVXC編碼器中 分別採用為0 · 9與0 . 4。W ( z )之脈衝響應經 DFT所得之離散振幅頻譜以向量β表示，以相 同之線性内插方式，1Η亦被轉換成固定維度向 量⑺(〇，ζ· = ι，2，..·，μ。在進行頻譜封包向量量 化時，搜尋最佳之碼字索引值/，是採用“最 小加權誤差評判準則”（M i n i m u m W e i g h t e d Distance)(步驟 13): 本紙張尺度適用中國國家標準（CNS)A4規格（210 X 297公釐） 573294 . A7 B7 五、發明說明（ r =arg min j Σ卜ω i=\ j = 0,1,. L 在 上式中，信 號r，及 分 別 向 量”及“加 權碼字 向. 量 ” ⑺ = r(F)(i)w(F)(〇, ( 4) = ^F)(/)w(F)(〇5 ( 5) 上述式（3 )經由推導可以得知，最佳之碼 字搜尋等同於找最大之正規化相關係數如 下： (請先閱讀背面之注意事項再填寫本頁) y* = arg max Σ<)(〜Γω = 〇, (6) 經濟部智慧財產局員工消費合作社印製 在式（4)，（5)需（L + 1)M 個乘法運算， 而式（6 )作最佳之碼字搜尋運算需 （2 Μ + 1 ) L 個乘法，L 個除法及2 M L 個加法。經由上述 方式搜尋得到最佳之碼字索引值/之後，該碼 字所需之最佳增益值儲存於增益表中） 可以依（7 )式求得： 、 g" 一x 在hvxc編碼器中利用封包增益量化表 G。，對此一最佳碼字增益值，c，採用Q (二5 ) 本紙張尺度適用中國國家標準（CNS)A4規格（210 X 297公釐） 573294 經濟部智慧財產局員工消費合作社印製 Α7 Β7 五、發明說明（） 位元純量量化。在解碼端固定維度封包向量 表不成· —々娜，（8) 其中，έ代表量化後之增益值。藉由編 碼端傳送之基週值，固定維度向量^F)可轉換 成可變維度向量f，利用△及隨機產生之相位 g = {A，i = l，2，."，P}，#*IDFT(InverseDiscrete Fourier Transform)方法，可得到合成之殘值 (激發）信號，此一諧波激發信號經 LPC(Linear Predictive Coding)合成濾波器 可輸出得到合成之語音信號。 由以上分析可知，Η V X C編碼器搜尋最佳 碼字所採用的誤差評判為最小加權誤差評判 準則；將頻譜封包向量轉換成固定維度之向 量，再進行向量量化。不僅由於使用維度轉 換而引入額外之轉換誤差，其頻譜封包向量 量化編碼演算法需要大量之計算。因為頻譜 封包向量量化編碼演算法之量化誤差對解碼 端的合成語音品質影響甚大，所以如何簡單 且有效的量化其封包向量，對該語音編解碼 器極為重要。 爰是之故，申請人有鑑於習知技術之缺 失，乃經悉心試驗與研究，發明出本案「可 變維度向量量化法」。 發明概述 8 本紙張尺度適用中國國家標準（CNS)A4規格（210 X 297公釐） -----------裝--------訂--------- (請先閱讀背面之注意事項再填寫本頁) 573294 A7 B7 五、發明說明（） (請先閱讀背面之注意事項再填寫本頁) 本案之主要目的係為利用人類耳朵之感 官特性，對聲音信號振幅之感知具備非線性 刻度及不同頻率具備不同頻率解析度之特 性，將各音框信號經過線性預估所得之殘值 信號的頻譜封包，以非線性的方式分割成固 定個數之頻帶，並計算各感官頻帶内所含諧 波頻譜之功率分貝值後，再作向量量化，以 大幅降低計算量及量化誤差。 經濟部智慧財產局員工消費合作社印製 本案之另一目的係為提供一種非線性分 頻向量量化的方法，用以將一固定Μ個維度 之頻譜封包向量進行向量量化，其步驟包含： (a )定義一 L * Ν個維度之對數刻度向量碼簿； (b )對該固定Μ個維度之頻譜封包向量進行 一等化處理後作非線性頻帶分割，進而得致 一 Ν個維度之頻譜封包向量；（c )將該Ν個維 度之頻譜封包向量取對數值以得致一第一目 標向量；（d )將該第一目標向量減去其平均值 以得致一第二目標向量；（e )將該對數刻度向 量碼薄其中之一 N個維度之碼字向量減去其 平均值以得致一基準碼字向量；（f )比較該第 二目標向量與該基準碼字向量之封包形 狀；（g )重覆步驟（e )至步驟（f )，以對該對數 刻度向量碼簿中之每一 N個維度之碼字向量 逐一進行比對；以及（h )根據該第二目標向 量與該基準碼字向量之比較結果而得致一位 於該對數刻度向量碼薄中之最佳碼字，進而 9 本紙張尺度適用中國國家標準（CNS)A4規格（210 X 297公釐） 經濟部智慧財產局員工消費合作社印製 573294 . Α7 Β7 五、發明說明（） 得致一量化過之頻譜封包向量。 根據上述構想，其中該固定Μ個維度之 頻譜封包向量係由一音框之Ρ個維度頻譜封 包向量經由維度轉換而成。 根據上述構想，其中該Ρ個維度頻譜封 包向量係為可變維度頻譜封包向量。 根據上述構想，其中該對數刻度向量碼 薄係由諧波及向量激發信號編碼器中之固定 維度向量碼薄以線性内差方式並取.其對數 值，預先轉換而成。 根據上述構想，其中該對數刻度向量碼 薄係具有L個碼字向量組合。 根據上述構想，其中該L個碼字向量組 合之任一係為Ν個維度之碼字向量。 根據上述構想，其中該等化處理係以一 預強調濾波器進行。 根據上述構想，若語音頻寬為4 Κ Η ζ，則 可分割成1 5個頻帶。 根據上述構想，其中頻帶1之對應中心 頻率為9 1 Η ζ，對應頻寬為9 1 Η ζ，並涵蓋了 該固定Μ個維度之頻譜封包向量之第1維 度。 根據上述構想，其中頻帶2之對應中心 頻率為1 8 2 Η ζ，對應頻寬為9 1 Η ζ，並涵蓋了 該固定Μ個維度之頻譜封包向量之第2維 度0 10 本紙張尺度適用中國國家標準（CNS)A4規格（210 X 297公釐） --------------------訂---------線 (請先閱讀背面之注意事項再填寫本頁) 573294 Α7 Β7 經濟部智慧財產局員工消費合作社印製 五、發明說明（） 根據上述構想，其中頻帶3之對應中心 頻率為2 7 3 Hz，對應頻寬為9 1 Η z，並涵蓋了 該固定Μ個維度之頻譜封包向量之第3維 度。 根據上述構想，其中頻帶4之對應中心 頻率為364Hz，對應頻寬為91Hz，並涵蓋了 該固定Μ個維度之頻譜封包向量之第4維 度。 根據上述構想，其中頻帶5之對應中心 頻率為5 Ο Ο Η ζ，對應頻寬為1 8 2 Η ζ，並涵蓋 了該固定Μ個維度之頻譜封包向量之第5至 第6維度。 根據上述構想，其中頻帶6之對應中心 頻率為68 1 Hz，對應頻寬為1 82Hz，並涵蓋 了該固定Μ個維度之頻譜封包向量之第7至 第8維度。 根據上述構想，其中頻帶7之對應中心 頻率為8 6 4 Η ζ，對應頻寬為1 8 2 Η ζ，並涵蓋 了該固定Μ個維度之頻譜封包向量之第9至 第1 0維度。 根據上述構想’其中頻帶8之對應中心 頻率為1 0 5 5 Η ζ，對應頻寬為1 8 2 Η ζ，並涵蓋 了該固定Μ個維度之頻譜封包向量之第1 1 至第1 2維度。 根據上述構想，其中頻帶9之對應中心 頻率為1 2 7 3 Η ζ，對應頻寬為2 7 3 Η ζ，並涵蓋 11 本紙張尺度適用中國國家標準（CNS)A4規格（210 X 297公釐） -----------裝--------訂--------- (請先閱讀背面之注意事項再填寫本頁) 經濟部智慧財產局員工消費合作社印製 573294 * Α7 Β7 五、發明說明（） 了該固定Μ個維度之頻譜封包向量之第1 3 至第1 5維度。 根據上述構想，其中頻帶1 0之對應中心 頻率為1 5 4 6 Η ζ，對應頻寬為2 7 3 Η ζ，並涵蓋 了該固定Μ個維度之頻譜封包向量之第1 6 至第1 8維度。 根據上述構想，其中頻帶1 1之對應中心 頻率為1 8 1 8 Η ζ，對應頻寬為2 7 3 Η ζ，並涵蓋 了該固定Μ個維度之頻譜封包向量之第1 9 至第2 1維度。 根據上述構想，其中頻帶1 2之對應中心 頻率為2136Hz，對應頻寬為364Hz，並涵蓋 了該固定Μ個維度之頻譜封包向量之第22 至第25維度。 根據上述構想，其中頻帶1 3之對應中心 頻率為2 5 4 6 Η ζ，對應頻寬為4 5 5 Η ζ，並涵蓋 了該固定Μ個維度之頻譜封包向量之第2 6 至第30維度。 根據上述構想，其中頻帶1 4之對應中心 頻率為3 0 4 6 Η ζ，對應頻寬為5 4 6 Η ζ，並涵蓋 了該固定Μ個維度之頻譜封包向量之第3 1 至第3 6維度。 根據上述構想，其中頻帶1 5之對應中心 頻率為3 6 3 6 Η ζ，對應頻寬為7 2 7 Η ζ，並涵蓋 了該固定Μ個維度之頻譜封包向量之第3 7 至第44維度。 12 本紙張尺度適用中國國家標準（CNS)A4規格（210 X 297公釐） -----------裝--------訂--------- (請先閱讀背面之注意事項再填寫本頁) 573294 . A7 B7 五、發明說明（） (請先閱讀背面之注意事項再填寫本頁) 根據上述構想，其中步驟（h )係利用一最 小頻譜距離準則（M i n i m u m S p e c t r a 1 Distance，MSD)來進行該最佳碼字之搜尋。 根據上述構想，其中步驟（h)更包含根據 該最佳碼字而求得一最佳碼字增益並將該最 佳碼字增益進行量化。 根據上述構想，其中係以Q位元純量量 化對該最佳碼字增益進行量化。 根據上述構想，其中步驟（h)更包含根據 該最佳碼字與該量化過之最佳碼字增益而得 致該量化過之.頻譜封包向量。 經濟部智慧財產局員工消費合作社印製 本案之又一目的係為提供一種非線性分 頻向量量化的方法，用以將一固定Μ個維度 之頻譜封包向量進行向量量化，其步驟包含： (a )定義一 L * Ν個維度之對數刻度向量碼薄； (b )對該固定Μ個維度之頻譜封包·向量進行 一等化處理後作非線性頻帶分割，進而得致 一 Ν個維度之頻譜封包向量；（.c )將該Ν個維 度之頻譜封包向量取對數值以得致一第一目 標向量；（d )將該第一目標向量減去其平均值 以得致一第二目標向量；（e )將該對數刻度向 量碼薄取得其中之一 N個維度之碼字向量以 得致一基準碼字向量；（f )比較該苐二目標向 量與該基準碼字向量之封包形狀；（g )重覆步 驟（e )至步驟（f )，以對該對數刻度向量碼薄 中之每一 N個維度之碼字向量逐一進行比 13 本紙張尺度適用中國國家標準（CNS)A4規格（210 X 297公釐） 573294 - A7 B7 五、發明說明（） (請先閱讀背面之注意事項再填寫本頁) 對；（h )根據該第二目標向量與該基準碼字向 量之比較結果而得致一位於該對數刻度向量 碼淳中之最佳碼午；以及（i )根據該最佳碼字 而求得一最佳碼字增益並將該最佳碼字增益 進行量化，進而得致一量化過之頻譜封包向 量。 根據上述構想，其中該固定Μ個維度之 頻譜封包向量係由一音框之Ρ個維度頻譜封 包向量經由維度轉換而成。 根據上述構想，其中該Ρ個維度頻譜封 包向量係為可.變維度頻譜封包向量。 根據上述構想，其中該對數刻度向量碼 薄係由譜波及向量激發信號編碼器中之固定 維度向量碼薄以線性内差方式變成Ν個為 度，並取其對數值後，預先扣除該向量之均 值轉換而成。 · 根據上述構想·其中該對數刻度向量碼 薄係具有L個碼字向量組合。 根據上述構想，其中該L個碼字向量組 合之任一係為Ν個維度之碼手向量。 經濟部智慧財產局員工消費合作社印製 根據上述構想，其中該等化處理係以一 預強調濾波器進行。 根據上述構想，若語音頻寬為4 Κ Η ζ，則 可分割成1 5個頻帶。 根據上述構想，其中頻帶1之對應中心 頻率為9 1 Η ζ，對應頻寬為9 1 Η ζ，並涵蓋了 14 本紙張尺度適用中國國家標準（CNS)A4規格（210 X 297 573294 Α7 Β7 經濟部智慧財產局員工消費合作社印製 五、發明說明（） 該固定Μ個維度之頻譜封包向量之第1維 度。 根據上述構想，其中頻帶2之對應中心 頻率為1 8 2 Η ζ，對應頻寬為9 1 Η ζ，並涵蓋了 該固定Μ個維度之頻譜封包向量之第2維 度。 根據上述構想，其中頻帶3之對應中心 頻率為2 7 3 Η ζ，對應頻寬為9 1 Η ζ，並涵蓋了 該固定Μ個維度之頻譜封包向量之第3維 度。 根據上述構想，其中頻帶4之對應中心 頻率為3 6 4 Η ζ，對應頻寬為9 1 Η ζ，並涵蓋了 該固定Μ個維度之頻譜封包向量之第4維 度。 根據上述構想，其中頻帶5之對應中心 頻率為5 Ο Ο Η ζ，對應頻寬為1 8 2 Η ·ζ，並涵蓋 了該固定Μ個維度之頻譜封包向量之第5至 第6維度。 根據上述構想，其中頻帶6之對應中心 頻率為6 8 1 Η ζ，對應頻寬為1 8 2 Η ζ，並涵蓋 了該固定Μ個維度之頻譜封包向量之第7至 第8維度。 根據上述構想，其中頻帶7之對應中心 頻率為8 6 4 Η ζ，對應頻寬為1 8 2 Η ζ，並涵蓋 了該固定Μ個維度之頻譜封包向量之第9至 第1 0維度。 15 本紙張尺度適用中國國家標準（CNS)A4規格（210 X 297公釐） --------------------訂--------- (請先閱讀背面之注意事項再填寫本頁) 573294 A7 B7五、發明說明（） 經濟部智慧財產局員工消費合作社印製 根據上述構想，其中頻帶8之對應中心 頻率為1 0 5 5 Η z，對應頻寬為1 8 2 Η z，並涵蓋 了該固定Μ個維度之頻譜封包向量之第1 1 至第1 2維度。 根據上述構想，其中頻帶9之對應中心 頻率為1 2 7 3 Η ζ，對應頻寬為2 7 3 Η ζ，並涵蓋 了該固定Μ個維度之頻譜封包向量之第13 至第1 5維度。 根據上述構想，其中頻帶1 0之對應中心 頻率為1 5 4 6 Hz，對應頻寬為2 7 3 Hz，並涵蓋 了該固定Μ個維度之頻譜封包向量之第1 6 至第1 8維度。 根據上述構想，其中頻帶1 1之對應中心 頻率為1 8 1 8 Η ζ，對應頻寬為2 7 3 Η ζ，並涵蓋 了該固定Μ個維度之頻譜封包向量之第1 9 至第2 1維度。 . 根據上述構想，其中頻帶1 2之對應中心 頻率為2 1 3 6 Η ζ，對應頻寬為3 6 4 Η ζ，並涵蓋 了該固定Μ個維度之頻譜封包向量之第2 2 至第25維度。 根據上述構想，其中頻帶.1 3之對應中心 頻率為2 5 4 6 Η ζ，對應頻寬為4 5 5 Η ζ，並涵蓋 了該固定Μ個維度之頻譜封包向量之第26 至第30維度。 根據上述構想，其中頻帶1 4之對應中心 頻率為3 0 4 6 Η ζ，對應頻寬為5 4 6 Η ζ，並涵蓋 (請先閱讀背面之注意事項再填寫本頁) ----訂---- 16 本紙張尺度適用中國國家標準（CNS)A4規格（210 X 297公釐） 573294 經濟部智慧財產局員工消費合作社印製 A7 _B7 五、發明說明（） 了該固定Μ個維度之頻譜封包向量之第3 1 至第3 6維度。 根據上述構想，其中頻帶1 5之對應中心 頻率為3 6 3 6 Η ζ，對應頻寬為7 2 7 Η ζ，並涵蓋 了該固定Μ個維度之頻譜封包向量之第3 7 至第44維度。 根據上述構想，其中步驟（h )係利用一最 小頻譜距離準貝1J ( M i n i m u m S p e c t r a 1 D i s t a n c e，M S D )來進行該最佳碼字之搜尋。 根據上述構想，其中係以Q位元純量量 化對該最佳碼.字增益進行量化。 根據上述構想，其中步驟（i )更包含根據 該最佳碼字與該量化過之最佳碼字增益而得 致該量化過之頻譜封包向量。 本案之再一目的係為提供一種非線性分 頻向量量化的方法，用以將一固定Μ個維度 之頻譜封包向量進行向量量化，其步驟包含： (a )定義一 L * Ν個維度之對數刻度向量碼簿； (b )對該固定Μ個維度之頻譜封包向量進行 一等化處理後作非線性頻帶分割，進而得致 一 Ν個維度之頻譜封包向量；.（c )將該Ν個維 度之頻譜封包向量取對數值以得致一第一目 標向量；（d )取該對數刻度向量碼薄其中之一 N個維度之碼字向量；（e )比較該第一目標向 量與該碼字向量之封包形狀；（f )重覆步驟（d ) 至步驟（e )，以對該對數刻度向量碼薄中之每 17 本紙張尺度適用中國國家標準（CNS)A4規格（210 X 297公釐） ---------------1----訂---------線 (請先閱讀背面之注意事項再填寫本頁) 573294 A7 B7 經1¾-部中夾¾卑而$工消贽合作社印^- 五、發明説明< ) 1 1 1 _ _ N 個 維 度 之 碼 字 向 量 逐 進 行 比 對 ;( g) 根 1 1 1 據 該1 第 一 g 標 向 量 與 該 碼 字 向 量 之 比 較 結 果 請 1 而 得 致 一 位 於 該 對 數 刻 度 向 量 碼 薄 中 之 r=7 取 佳 先 閱 1 | 石馬 字 ;以及（ h )根 據 該 最 佳 石馬 字 而 求 得 一 曰 取 佳 讀 背 1 碼 字 增 益 並 將 該 最 佳 碼 字 增 益 進 行 量 化 進 ¢7 之 注 1 I 而 得 致 一 曰 置 化 過 之 頻 譜 封 包 向 〇 意 事 1 1 根 據 上 述 構 想 其 中 該 固 定 Μ 個 維 度 之 項 再i 1 頻 譜 封 包 向 量 係 由 一 音 框 之 P 個 維 度 頻 譜 封 %} 本 裝 包 向 量 經 由 維 度 轉 換 而 成 〇 頁 1 I 根 據 上 述 構 想 其 中 該 P 個 維 度 頻 譜 封 1 I 包 向 量 係 為 可 變 維 度 頻 譜 封 包 向 量 0 1 1 根 據 上 述 構 想 其 中 該 對 數 刻 度 向 量 石馬 1 I 薄 係 由 諧 波 及 向 量 激 發 信 號 編 碼 器 中 之 固 定 1 訂 維 度 向 量 碼 薄 以 線 性 内 差 方 式 並 取 其 對 數 1 I 值 預 先 轉 換 而 成 0 1 I 根 據 上 述 構 想 其 中 該 對 數 刻 度 向 量 碼 1 薄 係 具 有 L 個 石馬 字 向 量 組 合 〇 1 I 根 據 上 述 構 想 其 中 該 L 個 石馬 字 向 量 組 I ά 合 之 任 一 係 為 N 個 維 度 之 石馬 字 向 量 0 1 I 根 據 上 述 構 想 其 中 該 等 化 處 理 係 以 一 1 I 預 強 調 慮 波 器 進 行 0 1 根 據 上 述 構 想 若 語 音 頻 見 為 4ΚΗζ， 則 1 I 可 分 割 成 1 5個頻帶 0 1 1 根 據 上 述 構 想 其 中 頻 、1 之 對 應 中 心 1 I 頻 率 為 9 1 Η z , 對應頻寬為S )1 Η z 並; '函 蓋 了 1 I 該 固 定 Μ :偃 1雖 [度 .之頻讀 1封包 丨向 1量 之第 1 維 1 1 1 18 1 1 1 本紙張尺度这用中國囤家標率（CNS ) Λ4規格（210X 297公釐） 573294 Α7 Β7 五、發明説明（） 度。 >艮據上述構想，其中頻帶2之對應中心 頻率為182Hz，對應頻寬為9 1 Η z，並涵蓋了 該固定Μ個維度之頻譜封包向量之第2維 度。 根據上述構想，其中頻帶3之對應中心 頻率為2 7 3 Η ζ，對應頻寬為9】Η ζ，並涵蓋了 該固定Μ個維度之頻譜封包向量之第3維 度。 根據上述構想，其中頻帶4之對應中心 頻率為3 6 4 Η ζ，對應頻寬為9 1 Η ζ，並涵蓋了 該固定Μ個維度之頻譜封包向量之第4維 度。 根據上述構想，其中頻帶5之對應中心 頻率為5 Ο Ο Η ζ，對應頻寬為1 8 2 Η ζ，並涵蓋 了該固定Μ個維度之頻譜封包向量之第5至 第6維度。 &淖部中央標準而只工消贽合作社印^- (請先閱讀背面之注意事項再填寫本頁) #1 根據上述構想，其中頻帶6之對應中心 頻率為6 8 1 Η ζ，對應頻寬為丨8 2 Η ζ，並涵蓋 了該固定Μ個維度之頻譜封包向量之第7至 第8維度。 根據上述構想，其中頻帶7之對應中心 頻率為8 6 4 Η ζ，對應頻寬為1 8 2 Η ζ，並涵蓋 了該固定Μ個維度之頻譜封包向量之第9至 第1 0維度。 根據上述構想，其中頻帶8之對應中心 19 本紙張尺度適川中國國家標嗥（CNS ) Λ4規格（210X 297公釐） 573294 A7 B7 五、發明説明（） 頻率為1 0 5 5 Η Z，對應頻寬為1 8 2 Η Z，並涵蓋 了該1固定Μ個維度之頻譜封包向量之第1 1 至第1 2維度。 根據上述構想，其中頻帶9之對應中心 頻率為1 2 7 3 Η ζ，對應頻寬為2 7 3 Η. ζ，並涵蓋 了該固定Μ個維度之頻譜封包向量之第1 3 至第1 5維度。 根據上述構想，其中頻帶1 0之對應中心 頻率為1 5 4 6 Η ζ，對應頻寬為2 7 3 Η ζ，並涵蓋 了該固定Μ個維度之頻譜封包向量之第1 6 至第1 8維度。 根據上述構想，其中頻帶1 1之對應中心 頻率為1818Hz，對應頻寬為2 7 3 Η ζ，並涵蓋 了該固定Μ個維度之頻譜封包向量之第1 9 至第2〗維度。 根據上述構想，其中頻帶1 2之對應中心 頻率為2 1 3 6 Η ζ，對應頻寬為3 6 4 Η ζ，並涵蓋 了該固定Μ個維度之頻譜封包向量之第22 至第25維度。 根據上述構想，其中頻帶1 3之對應中心 頻率為2 5 4 6 Η ζ，對應頻寬為4 5 5 Η ζ，並涵蓋 了該固定Μ個維度之頻譜封包向量之第26 至第30維度。 根據上述構想，其中頻帶14之對應中心 頻率為3 0 4 6 Η ζ，對應頻寬為5 4 6 Η ζ，並涵蓋 了該固定Μ個維度之頻譜封包向量之第3 1 20 紙張尺度適用中國國家標準（（:NS ) Λ4規格（210X 297公釐） 573294 Μ Β7五、發明説明（） 至第36維度。 1根據上述構想，其中頻帶1 5之對應中心 頻率為3 6 3 6 Η z，對應頻寬為7 2 7 Η z，並涵蓋 了該固定Μ個維度之頻譜封包向量之第3 7 至第44維度。 根據上述構想，其中步驟（h )係利用一最 小頻譜距離準貝1J ( M i n i m u m S p e c 11· a 1 D i s t a n c e，M S D )來進行該最佳碼字之搜尋。 根據上述構想，其中係以Q位元純量量 化對該最佳碼字增益進行量化。 根據上述構想，其中步驟（h )更包含根據 該最佳碼字與該量化過之最佳碼字增益而得 致該量化過之頻譜封包向量。 簡單圖式說明 第一圖：其係習用Η V X C語音編解碼器之頻譜 封包向量量化動作示意圖。 第二圖：其係本案之離散頻率與傳統連續性 自然頻率以非線性方式做頻帶分割後之維度 與頻帶對照比較圖。 第三圖：其係一般語音頻寬等化響度曲線 (Equal Loudness Curves)圖 〇 第四圖（a ):其係本案之離散頻率與傳統連續 性自然頻率以非線性方式做頻f分割後之頻 帶與頻寬對照比較圖。 第四圖（b ):其係本案之離散頻率與傳統連續 (誚先閱讀背面之注意事項再填寫本頁) -\口573294 A7 B7 V. Description of the invention () Field of invention (Please read the notes on the back before filling out this page) This case is a method of vector quantization ', especially a method of non-linear frequency division vector quantization. BACKGROUND OF THE INVENTION In the era of rapid growth of digital multimedia and the Internet, under the limited transmission bandwidth environment, more effective voice encoding and decoding methods must be developed to provide good communication service quality and Utilization rate does not cause any interruption. In the applications of artificial satellites and personal wireless mobile communications, how to meet the communication quality requirements at a lower bit rate is also a subject worthy of study. Printed by the Consumer Cooperative of the Intellectual Property Bureau of the Ministry of Economic Affairs. In response to the multi-purpose and multi-functional communication era in the 21st century, the International Moving Picture Experts Group (MPEG) has established the MPEG-4 audio protocol. In terms of speech coding technology, low-bit-rate (less than 4.8 kbps) "harmonic and vector excitation signal encoders" (H arm ο nic Vect 〇r Excitation Coding (HVXC)) are fixed at 4 kbps and 2 kbps The rate of the synthesized speech is higher than the current FS1016 encoding method with a bit rate of 4.8 kbps. The quality of the synthesized speech signal meets the requirement of "Communication Quality". In addition, the HVXC encoder also uses an adjustable type. Bit rate encoding and decoding architecture (bit rate 1.2k ~ 1.7kbps), and mention 3 on the decoding side. This paper size applies the Chinese National Standard (CNS) A4 specification (210 X 297 mm) 573294 A7 B7 Intellectual Property Bureau of the Ministry of Economic Affairs Printed by the employee consumer cooperative V. Invention description () Provides functions of variable pitch and variable speed. From this we can see that Η VXC voice editor The code system will gain attention and development in the near future, and may be widely used in communication and digital storage. Η The encoding method adopted by the VXC speech codec and the "multi-band excitation signal speech encoding method" (Multi- Band Excitation (MBE) coding method is similar, this kind of speech is wave coding (H arm ο niccoding) method, the voice signal in the voice signal (V oi ce d) sound box, using the harmonic spectrum amplitude vector quantization coding method; because The false number of speech harmonics is related to the base period of the speech frame, so vector quantization must be performed on the spectrum packet signal with time-varying dimensional characteristics. The coding algorithm not only consumes a lot of calculations, but also synthesizes the harmonic coding method. Speech quality depends on the quantization quality of the harmonic spectrum packet vector, so how to simply and effectively quantify its packet vector is an extremely important and valuable research topic for this coding method. In practice, the result of reducing the amount of calculation , Also helps in personal computer or DSP (Digital Signal Processor) micro-processing chip, Software or hardware implements the encoder and decoder on one chip at the same time, and performs real-time (rea 1-time) speech editing and decoding. HVXC speech encoder, voiced (voiced) or unvoiced (uniced) Different encoding methods are used for speech; traditional linear predictive coded encoding architecture is used for silent speech, and similar multi-band excitation letters are used for spoken speech ) -----------Install -------- Order --------- (Please read the precautions on the back before filling this page) Staff of the Intellectual Property Bureau of the Ministry of Economic Affairs Printed by the Consumer Cooperative 573294. A7 V. Description of the invention () The encoding method of the voice coding method. The encoding method is to first obtain the linear prediction parameters of each input sound box signal, and then pass the speech signal of the sound box through a residual value signal obtained by a linear prediction analysis filter, using a DFT of 2 5 6 points ( Discrete Fourier Transform) to obtain the spectrum and base period (Pi tch Lag) of its residual signal. For voice signals with sound, in order to transmit the residual signal spectrum, it is necessary to estimate the amplitude of the harmonic spectrum of each harmonic frequency band of the Fundamental Frequency, and The composed Spectral Envelope Vector (SEV) is used to find the best codeword and gain by the Spectral Envelope Vector Quantization (SEVQ) method and transmit the best codeword index value to the decoder. Since the dimension of the spectral packet vector is proportional to the size of the voice of the frame, the dimension of the spectral packet vector to be quantized by the spectral packet vector quantization is time-varying. (For a speech signal sampled at 8 k Η z, its spectral packet vector Dimension P ranges from 10 to 67). To solve this variable-dimensional vector quantization (VDVQ) problem, HVXC and speech encoders linearly interpolate the spectrum packet vector into a fixed dimension Μ (= 4 4) after the spectral packet shape vector f = ⑺ = 1,2, ··· M} (step 11), the vector codebook of r (= 8) bits is used for vector quantization. Please refer to the first figure, which is the conventional harmonic and vector language. 5 This paper size is applicable to China National Standard (CNS) A4 (210 X 297 mm) --------------- I ---- Order -------- line (please read the precautions on the back before filling this page) 573294 A7 B7 V. Description of the invention (Schematic diagram of vector quantization of spectrum packet of audio codec. Two each Codebook with K (= 2;? / 2 = 1 6) codeword vectors (first vector codebook ^^^ Γ, ^ 0, 1, ···, ^} Η and second vector codebook 15. In vector quantization, any codeword is composed of one codeword selected from the two codebooks. There are a total of L (= 2 5 6) possible codeword combinations; if / .. and h are used respectively Codebook P (F) and its cable bow I value, codeword vector of fixed dimension #, 7 · = feet 5% + A, as shown in equation (1): if%, / = 0, 1, ... ., 10 represents the linear prediction coefficient of the current coded sound frame, then the Perceptua 1 Weighting Filter 16 is as follows: 10Σα, / Z_ /! W {z) (Please read the precautions on the back first (Fill in this page again.) The company prints 10 10 \ u / Σ ^ 2 ^ '1' Σα ^ / = 0 / = 0 where the values of 6 and h are used in the HVXC encoder as 0 · 9 and 0.4 respectively. W (z The discrete amplitude spectrum of the impulse response obtained by DFT is represented by a vector β, and in the same linear interpolation manner, 1Η is also converted into a fixed-dimensional vector ⑺ (0, ζ · = ι, 2, .. ·, μ. In When performing spectral packet vector quantization, the search for the best codeword index / uses the “Minimum Weighted Error Judgment Criterion” (Step 13): This paper scale applies the Chinese National Standard (CNS) A4 specification (210 X 297 mm) 573294. A7 B7 V. Description of the invention (r = arg min j Σωω i = \ j = 0,1 ,. L In the above formula, the signal r, and its respective vector "and" weight Code direction. Quantities ”⑺ = r (F) (i) w (F) (〇, (4) = ^ F) (/) w (F) (〇5 (5) The above formula (3) can be derived by It is learned that the best codeword search is equivalent to finding the largest normalized correlation coefficient as follows: (Please read the precautions on the back before filling this page) y * = arg max Σ <) (~ Γω = 〇, (6) Ministry of Economy Wisdom The Property Cooperative's Consumer Cooperative printed in formula (4), (5) requires (L + 1) M multiplication operations, and formula (6) for the best codeword search operation requires (2 MU + 1) L multiplications , L divisions and 2 ML additions. After searching through the above method to obtain the best codeword index value /, the optimal gain value required for the codeword is stored in the gain table.) It can be obtained according to formula (7):, g " -x in the hvxc encoder Use the packet gain quantization table G. For this, the best codeword gain value, c, uses Q (2 5). This paper size applies the Chinese National Standard (CNS) A4 specification (210 X 297 mm). 573294 Printed by the Consumers ’Cooperative of Intellectual Property Bureau of the Ministry of Economic Affairs Α7 Β7 5. Description of the invention () Bit scalar quantization. The fixed-dimensional packet vector on the decoding side does not make it 々 々, (8) where έ represents the gain value after quantization. The fixed-dimensional vector ^ F) can be converted into a variable-dimensional vector f by using the base-period value transmitted from the encoding end. Using △ and a randomly generated phase g = {A, i = 1, 2,. &Quot;, P}, # * IDFT (InverseDiscrete Fourier Transform) method can obtain the synthesized residual value (excitation) signal. This harmonic excitation signal can be output through the Linear Predictive Coding (LPC) synthesis filter to obtain the synthesized speech signal. From the above analysis, it can be known that the error judgment used by the Η V X C encoder to search for the best codeword is the minimum weighted error judgment criterion; the spectrum packet vector is converted into a vector of a fixed dimension, and then vector quantization is performed. Not only is the extra conversion error introduced due to the use of dimensional conversion, the spectral packet vector quantization coding algorithm requires a lot of calculations. Because the quantization error of the spectral packet vector quantization coding algorithm has a great impact on the quality of the synthesized speech on the decoder side, how to simply and effectively quantify its packet vector is extremely important for this speech codec. For this reason, the applicant has invented the “variable dimensional vector quantification method” in this case after careful testing and research in view of the lack of known technology. Summary of the Invention 8 This paper size is applicable to China National Standard (CNS) A4 specification (210 X 297 mm) ----------- installation -------- order ------- -(Please read the notes on the back before filling this page) 573294 A7 B7 V. Description of the invention () (Please read the notes on the back before filling this page) The main purpose of this case is to use the sensory characteristics of the human ear, The perception of the amplitude of the sound signal has a characteristic of non-linear scale and different frequencies with different frequency resolution. Frequency band, and after calculating the power decibel value of the harmonic spectrum contained in each sensory frequency band, vector quantization is performed to greatly reduce the calculation amount and quantization error. Another objective of the case printed by the Employees' Cooperatives of the Intellectual Property Bureau of the Ministry of Economic Affairs is to provide a method of non-linear frequency division vector quantization for vector quantization of a fixed M-dimensional spectrum envelope vector. The steps include: (a ) Define a logarithmic scale vector codebook of L * N dimensions; (b) perform equalization processing on the spectrum packet vectors of fixed M dimensions and perform non-linear band division, thereby obtaining spectrum packets of N dimensions Vector; (c) take the logarithmic values of the spectrum packet vectors of the N dimensions to obtain a first target vector; (d) subtract the average of the first target vector to obtain a second target vector; ( e) Subtract the average of the codeword vectors in one of the N dimensions of the logarithmic scale vector codebook to obtain a reference codeword vector; (f) compare the packet of the second target vector with the reference codeword vector Shape; (g) repeating steps (e) to (f) to compare the codeword vectors of each N dimensions in the logarithmic scale vector codebook one by one; and (h) according to the second goal Vector with the benchmark The result of the comparison of the word vectors resulted in the best codeword in the logarithmic scale vector codebook, and 9 paper sizes are applicable to the Chinese National Standard (CNS) A4 specification (210 X 297 mm). Employees of the Intellectual Property Bureau of the Ministry of Economic Affairs Printed by the Consumer Cooperative 573294. Α7 Β7 V. Description of the Invention () A quantized spectrum packet vector is obtained. According to the above-mentioned concept, the fixed M-dimensional spectrum packet vector is converted from the P-dimensional spectrum packet vector of a sound frame through dimensions. According to the above concept, the P-dimensional spectral packet vectors are variable-dimensional spectral packet vectors. According to the above-mentioned concept, the logarithmic scale vector codebook is obtained by pre-converting the logarithmic value of the fixed-dimensional vector codebook in the harmonic and vector excitation signal encoder in a linear internal difference manner. According to the above concept, the logarithmic scale vector codebook has L codeword vector combinations. According to the above idea, any one of the L codeword vector combinations is a codeword vector of N dimensions. According to the above idea, the equalization process is performed by a pre-emphasis filter. According to the above idea, if the speech and audio width is 4K Η ζ, it can be divided into 15 frequency bands. According to the above concept, the corresponding center frequency of the frequency band 1 is 9 1 Η ζ, and the corresponding bandwidth is 9 1 Η ζ, and it covers the first dimension of the spectrum packet vector of the fixed M dimensions. According to the above idea, the corresponding center frequency of Band 2 is 1 8 2 Η ζ, and the corresponding bandwidth is 9 1 Η ζ, and it covers the second dimension of the spectrum envelope vector of the fixed M dimensions. 0 10 This paper scale applies to China National Standard (CNS) A4 Specification (210 X 297 mm) -------------------- Order --------- Line (Please read the back first Please note this page, please fill in this page again) 573294 Α7 Β7 Printed by the Consumer Cooperatives of the Intellectual Property Bureau of the Ministry of Economic Affairs 5. Description of the invention () According to the above idea, the corresponding center frequency of Band 3 is 2 7 3 Hz and the corresponding bandwidth is 9 1 Η z, and covers the third dimension of the spectrum packet vector of the fixed M dimensions. According to the above idea, the corresponding center frequency of the band 4 is 364 Hz, and the corresponding bandwidth is 91 Hz, and it covers the fourth dimension of the spectrum packet vector of the fixed M dimensions. According to the above idea, the corresponding center frequency of the frequency band 5 is 5 〇 Η ζ ζ, and the corresponding bandwidth is 1 8 2 Η ζ, and covers the 5th to 6th dimensions of the fixed M-dimensional spectrum envelope vector. According to the above idea, the corresponding center frequency of the band 6 is 68 1 Hz, and the corresponding bandwidth is 1 82 Hz, and covers the seventh to eighth dimensions of the spectrum packet vector of the fixed M dimensions. According to the above idea, the corresponding center frequency of the band 7 is 8 6 4 Η ζ, and the corresponding bandwidth is 1 8 2 Η ζ, and covers the 9th to 10th dimensions of the spectrum packet vector of the fixed M dimensions. According to the above conception, where the corresponding center frequency of band 8 is 1 0 5 5 Η ζ and the corresponding bandwidth is 1 8 2 Η ζ, and covers the 1st to 12th dimensions of the spectrum packet vector of the fixed M dimensions . According to the above idea, the corresponding center frequency of the band 9 is 1 2 7 3 Η ζ, the corresponding bandwidth is 2 7 3 Η ζ, and it covers 11 paper sizes. The Chinese national standard (CNS) A4 specification (210 X 297 mm) is applicable. ) ----------- Installation -------- Order --------- (Please read the notes on the back before filling this page) Employees of the Intellectual Property Bureau of the Ministry of Economic Affairs Printed by the consumer cooperative 573294 * Α7 Β7 V. Description of the invention () The 13th to 15th dimensions of the spectrum packet vector of the fixed M dimensions are described. According to the above idea, the corresponding center frequency of the band 10 is 1 5 4 6 Η ζ, and the corresponding bandwidth is 2 7 3 Η ζ, and covers the 16th to 18th of the spectrum packet vector of the fixed M dimensions. Dimensions. According to the above idea, the corresponding center frequency of the frequency band 1 1 is 1 8 1 8 Η ζ, and the corresponding bandwidth is 2 7 3 Η ζ, and covers the 19th to the 21st of the spectrum packet vector of the fixed M dimensions. Dimensions. According to the above idea, the corresponding center frequency of the frequency band 12 is 2136 Hz and the corresponding bandwidth is 364 Hz, and covers the 22nd to 25th dimensions of the spectrum packet vector of the fixed M dimensions. According to the above concept, the corresponding center frequency of the band 13 is 2 5 4 6 Η ζ, and the corresponding bandwidth is 4 5 5 Η ζ, and covers the 26th to 30th dimensions of the spectrum packet vector of the fixed M dimensions . According to the above concept, the corresponding center frequency of the band 14 is 3 0 4 6 Η ζ, and the corresponding bandwidth is 5 4 6 Η ζ, and covers the 3rd to 3rd of the spectrum packet vector of the fixed M dimensions. Dimensions. According to the above idea, the corresponding center frequency of the frequency band 15 is 3 6 3 6 Η ζ, and the corresponding bandwidth is 7 2 7 Η ζ, and covers the 37th to 44th dimensions of the spectrum packet vector of the fixed M dimensions . 12 This paper size applies to China National Standard (CNS) A4 specification (210 X 297 mm) ----------- installation -------- order --------- (Please read the notes on the back before filling this page) 573294. A7 B7 V. Description of the invention () (Please read the notes on the back before filling this page) According to the above idea, step (h) is to use a minimum spectrum The distance criterion (Minimum Spectra 1 Distance, MSD) is used to search for the best codeword. According to the above concept, step (h) further comprises obtaining an optimal codeword gain based on the optimal codeword and quantizing the optimal codeword gain. According to the above concept, the optimal codeword gain is quantized by quantizing the Q bits. According to the above concept, step (h) further includes obtaining the quantized .spectral packet vector according to the best codeword and the quantized best codeword gain. Another objective of the case printed by the Employees' Cooperative of the Intellectual Property Bureau of the Ministry of Economic Affairs is to provide a method of non-linear frequency division vector quantization for vector quantization of a fixed M-dimensional spectrum envelope vector. The steps include: (a ) Define a logarithmic scale vector codebook of L * N dimensions; (b) Unify the frequency spectrum packet and vector of fixed M dimensions and perform non-linear frequency band division to obtain a spectrum of N dimensions Packet vector; (.c) take logarithmic values of the spectrum packet vectors of N dimensions to obtain a first target vector; (d) subtract the average of the first target vector to obtain a second target vector (E) obtaining one of the N codeword vectors from the logarithmic scale vector codebook to obtain a reference codeword vector; (f) comparing the packet shape of the second target vector and the reference codeword vector; (G) Repeat steps (e) to (f) to compare the codeword vectors of each N dimensions in the logarithmic scale vector codebook one by one. 13 This paper size applies the Chinese National Standard (CNS) A4 specification. (210 X 29 7mm) 573294-A7 B7 V. Description of the invention () (Please read the notes on the back before filling this page) Yes; (h) It is based on the comparison result between the second target vector and the reference codeword vector An optimal code day located in the logarithmic scale vector code; and (i) obtaining an optimal codeword gain based on the optimal codeword and quantizing the optimal codeword gain, thereby obtaining a quantization Passed spectrum packet vector. According to the above-mentioned concept, the fixed M-dimensional spectrum packet vector is converted from the P-dimensional spectrum packet vector of a sound frame through dimensions. According to the above concept, the P-dimensional spectral packet vectors are variable-dimensional spectral packet vectors. According to the above-mentioned concept, the logarithmic scale vector codebook is changed from the fixed-dimensional vector codebook of the spectral wave and vector excitation signal encoder to N degrees in a linear internal difference manner, and after taking its logarithmic value, the vector is deducted in advance. The mean is converted. According to the above idea, wherein the logarithmic scale vector code sheet has L codeword vector combinations. According to the above idea, any one of the L codeword vector combinations is a code hand vector of N dimensions. Printed by the Consumer Cooperative of the Intellectual Property Bureau of the Ministry of Economic Affairs According to the above idea, the chemical treatment is performed by a pre-emphasis filter. According to the above idea, if the speech and audio width is 4K Η ζ, it can be divided into 15 frequency bands. According to the above idea, the corresponding center frequency of Band 1 is 9 1 Η ζ, and the corresponding bandwidth is 9 1 Η ζ, and it covers 14 paper standards applicable to the Chinese National Standard (CNS) A4 specification (210 X 297 573294 Α7 Β7 Economy Printed by the Consumer Cooperatives of the Ministry of Intellectual Property Bureau. 5. Description of the invention () The first dimension of the spectrum envelope vector of the fixed M dimensions. According to the above idea, the corresponding center frequency of Band 2 is 1 8 2 Η ζ, corresponding to the bandwidth. Is 9 1 Η ζ, and covers the second dimension of the spectrum packet vector of the fixed M dimensions. According to the above concept, the corresponding center frequency of band 3 is 2 7 3 Η ζ, and the corresponding bandwidth is 9 1 Η ζ, It also covers the third dimension of the spectrum packet vector of the fixed M dimensions. According to the above idea, the corresponding center frequency of the frequency band 4 is 3 6 4 Η ζ and the corresponding bandwidth is 9 1 Η ζ, and the fixed M is covered. The 4th dimension of the spectrum packet vector of each dimension. According to the above idea, the corresponding center frequency of the frequency band 5 is 5 Ο Η Η ζ, the corresponding bandwidth is 1 8 2 Η · ζ, and the spectrum of the fixed M dimensions is covered. seal 5th to 6th dimensions of the vector. According to the above concept, the corresponding center frequency of the frequency band 6 is 6 8 1 Η ζ and the corresponding bandwidth is 1 8 2 Η ζ, and it covers the fixed M dimensions of the spectrum envelope vector. 7th to 8th dimensions. According to the above idea, the corresponding center frequency of the band 7 is 8 6 4 Η ζ, and the corresponding bandwidth is 1 8 2 Η ζ, and it covers the ninth of the spectrum packet vector of the fixed M dimensions. To the 10th dimension. 15 This paper size applies to China National Standard (CNS) A4 (210 X 297 mm) -------------------- Order --- ------ (Please read the notes on the back before filling this page) 573294 A7 B7 V. Description of Invention () Printed by the Consumer Cooperatives of the Intellectual Property Bureau of the Ministry of Economic Affairs According to the above concept, the corresponding center frequency of Band 8 is 1 0 5 5 Η z, the corresponding bandwidth is 1 8 2 Η z, and covers the 1st to 12th dimensions of the spectrum packet vector of the fixed M dimensions. According to the above concept, the corresponding center frequency of the frequency band 9 Is 1 2 7 3 Η ζ, the corresponding bandwidth is 2 7 3 Η ζ, and covers the thirteenth of the spectrum packet vector of the fixed M dimensions To the 15th dimension. According to the above idea, the corresponding center frequency of the frequency band 10 is 15 4 6 Hz and the corresponding bandwidth is 2 73 Hz, and it covers the 16th of the spectrum packet vector of the fixed M dimensions. To the eighteenth dimension. According to the above concept, the corresponding center frequency of the frequency band 11 is 1 8 1 8 Η ζ, and the corresponding bandwidth is 2 7 3 Η ζ, and covers the first of the spectrum packet vectors of the fixed M dimensions. 1 9 to 21st dimension. According to the above idea, the corresponding center frequency of the band 12 is 2 1 3 6 Η ζ, and the corresponding bandwidth is 3 6 4 Η ζ, and covers the 2nd to 25th of the spectrum packet vector of the fixed M dimensions. Dimensions. According to the above idea, the corresponding center frequency of the frequency band .1 3 is 2 5 4 6 Η ζ, and the corresponding bandwidth is 4 5 5 Η ζ, and covers the 26th to 30th dimensions of the spectrum packet vector of the fixed M dimensions. . According to the above idea, the corresponding center frequency of the band 14 is 3 0 4 6 Η ζ, and the corresponding bandwidth is 5 4 6 Η ζ, and it is covered (please read the precautions on the back before filling this page) ---- order ---- 16 This paper size applies the Chinese National Standard (CNS) A4 specification (210 X 297 mm) 573294 Printed by the Consumers' Cooperatives of the Intellectual Property Bureau of the Ministry of Economic Affairs A7 _B7 V. Description of the invention () The fixed M dimensions The 31st to 36th dimensions of the spectrum packet vector. According to the above idea, the corresponding center frequency of the frequency band 15 is 3 6 3 6 Η ζ, and the corresponding bandwidth is 7 2 7 Η ζ, and covers the 37th to 44th dimensions of the spectrum packet vector of the fixed M dimensions . According to the above concept, the step (h) is to search for the best codeword by using a minimum spectral distance quasi-beam 1J (M i n m m p ec t r a 1 D i s t c n c, M S D). According to the above concept, the optimal code and word gain is quantized by scalar Q-bit quantization. According to the above concept, step (i) further comprises obtaining the quantized spectrum packet vector according to the best codeword and the quantized best codeword gain. Another purpose of this case is to provide a method of non-linear frequency division vector quantization for vector quantization of a spectrum packet vector of fixed M dimensions. The steps include: (a) defining a logarithm of L * N dimensions Scale vector codebook; (b) performing equalization processing on the spectrum packet vectors of fixed M dimensions and dividing them into non-linear bands to obtain a spectrum packet vector of N dimensions; (c) the N number of Take the logarithmic value of the spectral envelope vector of the dimension to obtain a first target vector; (d) take a codeword vector of N dimensions in one of the logarithmic scale vector codebooks; (e) compare the first target vector with the code The shape of the packet of the word vector; (f) Repeat steps (d) to (e) to apply the Chinese National Standard (CNS) A4 specification (210 X 297 mm) to every 17 paper sizes in the logarithmic scale vector codebook Li) --------------- 1 ---- order --------- line (please read the notes on the back before filling this page) 573294 A7 B7 1¾-In the middle of the ministry ¾ and the labor and consumption cooperative seal ^-5. Description of the invention <) 1 1 1 _ _ N dimensions The codeword vectors are compared one by one; (g) root 1 1 1 According to the comparison result between the first g-vector of the 1 and the codeword vector, please take 1 to get r = 7 in the logarithmic scale vector codebook. Take the best first read 1 | Shima; and (h) find the best Shima word based on the best read back 1 codeword gain and quantize the best codeword gain into Note 1 of ¢ 7 I get the result of the set spectrum packet orientation. 0 Meaning 1 1 According to the above concept, the fixed M dimensions of the item are then i 1 The spectrum packet vector is the spectrum of the P dimensions of a frame.%} The packet vector is converted into dimensions. Page 1 I According to the above concept, the P dimension spectrum envelope 1 I The packet vector is a variable-dimensional spectrum envelope vector 0 1 1 According to the above concept, the logarithmic scale vector Shima 1 I thin system Encoding from harmonic and vector excitation signals The fixed 1-dimensional vector codebook in the device is converted into a linear internal difference and its logarithmic 1 value is converted into 0 1 I according to the above concept. The logarithmic scale vector code 1 thin system has L stone horse vector combinations. 1 I According to the above-mentioned concept, the L stone horse character vector group I is either a N-dimensional stone horse word vector. 0 1 I According to the above idea, the equalization process is a 1 I pre-emphasis on the consideration wave. The device performs 0 1 according to the above concept. If the speech and audio are 4KΗζ, then 1 I can be divided into 15 frequency bands. 0 1 1 According to the above concept, the corresponding center of the intermediate frequency and 1 1 I frequency is 9 1 Η z, and the corresponding bandwidth is S) 1 Η z ;; 'The letter covers 1 I The fixed M: 偃 1 Although [degree. Frequency reads 1 packet 丨 to the 1st dimension 1 1 1 1 18 1 1 1 Standard rate (CNS) Λ4 specification (210X 297 male ) 573294 Α7 Β7 V. invention is described in () degrees. > According to the above idea, the corresponding center frequency of Band 2 is 182 Hz and the corresponding bandwidth is 9 1 Η z, and it covers the second dimension of the spectrum packet vector of the fixed M dimensions. According to the above concept, the corresponding center frequency of the band 3 is 2 7 3 Η ζ, and the corresponding bandwidth is 9] ζ ζ, and it covers the third dimension of the spectrum packet vector of the fixed M dimensions. According to the above idea, the corresponding center frequency of the frequency band 4 is 3 6 4 Η ζ, and the corresponding bandwidth is 9 1 Η ζ, and covers the fourth dimension of the spectrum packet vector of the fixed M dimensions. According to the above idea, the corresponding center frequency of the frequency band 5 is 5 〇 Η ζ ζ, and the corresponding bandwidth is 1 8 2 Η ζ, and covers the 5th to 6th dimensions of the fixed M-dimensional spectrum envelope vector. & Central Standards of the Ministry of Industry, and only printed by cooperatives ^-(Please read the notes on the back before filling this page) # 1 According to the above idea, the corresponding center frequency of Band 6 is 6 8 1 Η ζ, the corresponding frequency The width is 8 2 Η ζ and covers the seventh to eighth dimensions of the spectrum packet vector of the fixed M dimensions. According to the above idea, the corresponding center frequency of the band 7 is 8 6 4 Η ζ, and the corresponding bandwidth is 1 8 2 Η ζ, and covers the 9th to 10th dimensions of the spectrum packet vector of the fixed M dimensions. According to the above idea, the corresponding center of the band 8 is 19 paper sizes, which is suitable for the Chinese National Standard (CNS) Λ4 specification (210X 297 mm) 573294 A7 B7 V. Description of the invention () The frequency is 1 0 5 5 Η Z, corresponding The bandwidth is 1 8 2 Η Z and covers the 1st to 12th dimensions of the spectrum packet vector of the 1 fixed M dimensions. According to the above idea, the corresponding center frequency of the frequency band 9 is 1 2 7 3 Η ζ, and the corresponding bandwidth is 2 7 3 Η. Ζ, and covers the 1st to 1st 5th of the spectrum packet vector of the fixed M dimensions. Dimensions. According to the above idea, the corresponding center frequency of the band 10 is 1 5 4 6 Η ζ, and the corresponding bandwidth is 2 7 3 Η ζ, and covers the 16th to 18th of the spectrum packet vector of the fixed M dimensions. Dimensions. According to the above idea, the corresponding center frequency of the band 11 is 1818 Hz, and the corresponding bandwidth is 2 7 3 Η ζ, and covers the 19th to 2nd dimensions of the spectrum packet vector of the fixed M dimensions. According to the above idea, the corresponding center frequency of the frequency band 12 is 2 1 3 6 Η ζ, and the corresponding bandwidth is 3 6 4 Η ζ, and covers the 22nd to 25th dimensions of the spectrum packet vector of the fixed M dimensions. According to the above concept, the corresponding center frequency of the band 13 is 2 5 4 6 Η ζ, and the corresponding bandwidth is 4 5 5 Η ζ, and covers the 26th to 30th dimensions of the fixed M-dimensional spectrum packet vector. According to the above idea, the corresponding center frequency of band 14 is 3 0 4 6 Η ζ, the corresponding bandwidth is 5 4 6 Η ζ, and the 3rd 20th paper scale covering the fixed M dimensions of the spectrum envelope vector is applicable to China National standard ((: NS) Λ4 specification (210X 297 mm) 573294 Μ B7 V. Description of the invention () to the 36th dimension. 1 According to the above concept, the corresponding center frequency of the band 15 is 3 6 3 6 Η z, The corresponding bandwidth is 7 2 7 Η z and covers the 37th to 44th dimensions of the spectrum packet vector of the fixed M dimensions. According to the above concept, step (h) uses a minimum spectral distance of 1J ( Mimim Spec 11.11 Distance (MSD) to perform the search for the best codeword. According to the above idea, the best codeword gain is quantized by quantizing Q bits scalar. According to the above idea, The step (h) further includes obtaining the quantized spectrum packet vector according to the best codeword and the quantized best codeword gain. A simple diagram illustrates the first picture: it is a conventional ΗVXC speech codec Spectrum packet Schematic diagram of quantitative quantization action. Figure 2: It is a comparison chart between the discrete frequency and the traditional continuous natural frequency in a non-linear manner after the frequency band division. The third figure: It is the generalized loudness of general speech and audio. Figure (Equal Loudness Curves) Figure 4 (a): This is the case of the discrete frequency and the traditional continuous natural frequency in a non-linear manner after the frequency f-segment frequency band and bandwidth comparison chart. The fourth figure (b ): It is the discrete frequency and traditional continuity of this case (诮 Please read the notes on the back before filling this page)-\ 口
•I 21 本紙張尺度適川中國S家標準（CNS ) Λ4規格（210X 297公釐） 573294 f A7 B7 五、發明説明（） 性自然頻率以非線性方式做頻帶分割後之頻 帶與1中心頻率對照比較圖。 第五圖：其係本案一較佳實施例之動作示意 圖。 第六圖（a):其係一音框語音信號之信號波形 圖。 第六圖（b):其係一音框語音信號之頻譜圖。 第六圖（c ):其係以本案之方法與以習用 HVXC向量量化法對一音框語音信號進行量 化後之頻譜封包比較圖。 第七圖（a):其係一段語音信號之信號波形 圖。 第七圖（b )··其係一段語音信號之基週變化曲 線圖。 第七圖（c):其係以本案之方法與以習用 HVXC向量量化法對一段語音信號進行量化 後之頻譜失真比較圖。 (讀先閱讀背面之注意事項再填寫本頁) 薄 碼 薄器量 碼波向 量濾度 表向調刻 益二強數 增第預對 2 5 18 1Χ IX πο ΠΟ 第 向 加向 覺字 聽碼 器表 薄波值 矹濾均V旦里 頻 型 新 明一 說出 例提 施案 實本 佳 較 方 化 量 量 向 包 圭 2 2 本紙張尺度適用中國國家標準（CNS ) Λ4規格（210X29*7公釐） 573294 ^ A7 B7 經濟部智慧財產局員工消費合作社印製 五、發明說明（） 式，根據人類耳朵對聲音信號頻率及振幅之 感知，具備非線性刻度（Ν ο η 1 i n e a r S c a 1 e )之特 性，來建構向量量化之架構。以下，首先說 明如何利用人耳非線性頻率解析度之特性， 將頻譜封包向量作非線性頻帶分割，再說明 以何種向量量化方式，以達到降低計算複雜 度、減少量化誤差並提高合成語音的品質目 的。最後以客觀之頻譜失真量測，評估本發 明所提方法之性能，並以理論分析其計算複 雜度。 (一）、非線性頻帶分割 由於人耳聽覺對頻率之解析度有限，且 對高低頻率亦有不同之解析度（對低頻率信 號之頻率解析度較高），本案以非線性方式做 頻帶分割成若干個感官頻帶，如第二圖所 示。除此之外，人耳對聲音信號之振幅會因 不同頻率位置而有不同之感受度，如第三圖 所示，耳朵感受之音壓位準（Sound Pressure Level, SPL)，以分貝為單位。對語音通訊頻 寬，人耳對頻率2kHz以上之信號的感受度較 強，此一現象，可以用一預強調濾波器 H(z) ，//⑺=(2.6 + ^)/(1.6 + 2-1)，將不同頻率之語音 信號作等化處理，使得相同之信號能量，在 不同之頻率有相同之響度。 利用上述之人耳感官模型，對信號作編 23 本紙張尺度適用中國國家標準（CNS)A4規格（210 x 297公釐） ---------------------訂---------線I赢 (請先閱讀背面之注意事項再填寫本頁) -ϋ I — ϋ n n n I ^ ϋ ϋ ϋ n ϋ n I ·1 ^ ϋ n ϋ - 573294 ^ A7 _B7 五、發明說明（） (請先閱讀背面之注意事項再填寫本頁) 碼時，先計算各個頻帶之信號能量，並換算 成人耳主觀感受之響度位準（Loudness level) 再作編碼，已被證實能有效的編碼音樂信 號。而自然頻率 /(Hz) 映對成巴克（Bark) 刻度 6 的方程式可描述為： 0 = 681^-1^ / = 0〜从仲）（9) 對每一 Bark頻帶，其-3dB頻寬為一個 Bark刻度，每一感官頻帶上下限、中心頻率 及頻寬可分別表示如下： fl(n) = 600sinh(^^) ，（10) 6 />) = 60〇Sinh(^^)，（11) 〇 /c ⑻= 600sinh(昏)，（12) 〇 cb\\{n) = fu{n)- ft{n) (13) 由文獻研究得知，對頻寬4kHz之語音頻 寬，可分割成N個（=1 5) Bark頻帶，如表（一） 所示。 經濟部智慧財產局員工消費合作社印製 24 本紙張尺度適用中國國家標準（CNS)A4規格（210 X 297公釐） 573294 ‘ A7 B7 經濟部智慧財產局員工消費合作社印製 五、發明說明（） 表(一）：自然頻率映對Bark之參數表(單位:Hz)• I 21 This paper is suitable for Sichuan S Standard (CNS) Λ4 specification (210X 297mm) 573294 f A7 B7 V. Description of the invention () The natural frequency is divided into non-linear bands and a center frequency Contrast comparison chart. Fifth figure: It is a schematic diagram of the operation of a preferred embodiment of the present case. Figure 6 (a): It is a signal waveform diagram of a sound frame voice signal. The sixth figure (b): it is a spectrum diagram of a sound frame speech signal. Figure 6 (c): This is a comparison chart of the spectral packet after quantizing a sound frame voice signal using the method of this case and the conventional HVXC vector quantization method. The seventh figure (a): it is a signal waveform diagram of a section of speech signal. The seventh diagram (b) is a curve diagram of the base period of a segment of speech signal. Figure 7 (c): This is a comparison chart of spectral distortion after quantizing a speech signal using the method of this case and the conventional HVXC vector quantization method. (Read the precautions on the back before you fill in this page.) The thin-code thin-quantity code wave vector filter table is added to the tuning gain second strong number. The pre-pair is 2 5 18 1 × IX πο ΠΟ. Table thin wave values are filtered by V-density frequency. Xin Mingyi exemplifies the actual implementation of the case. The amount of parity is better. 2 2 This paper size applies to Chinese National Standard (CNS) Λ4 specification (210X29 * 7). (%) 573294 ^ A7 B7 Printed by the Consumer Cooperatives of the Intellectual Property Bureau of the Ministry of Economic Affairs 5. Description of the invention () formula, based on the human ear's perception of the frequency and amplitude of sound signals, with a non-linear scale (N ο η 1 inear S ca 1 e ) To construct a vector quantization architecture. In the following, we first explain how to use the characteristics of the non-linear frequency resolution of the human ear to divide the spectral packet vector into non-linear bands, and then explain which vector quantization method is used to reduce the computational complexity, reduce the quantization error, and improve the quality of the synthesized speech. Quality purpose. Finally, objective spectral distortion measurement is used to evaluate the performance of the method proposed in the present invention, and its theoretical complexity is calculated by theoretical analysis. (I) Non-linear frequency band segmentation Because the human hearing has limited resolution of frequency and different resolutions for high and low frequencies (higher frequency resolution for low-frequency signals), this case uses a non-linear manner to perform frequency band segmentation. Into several sensory frequency bands, as shown in the second figure. In addition, the amplitude of the human ear's sound signal will have different sensitivity due to different frequency positions. As shown in the third figure, the sound pressure level (SPL) felt by the ear is in decibels. . For the voice communication bandwidth, the human ear is more sensitive to signals above 2kHz. This phenomenon can be pre-emphasized by H (z), //⑺=(2.6 + ^) / (1.6 + 2 -1) equalize the speech signals of different frequencies so that the same signal energy has the same loudness at different frequencies. Using the above-mentioned human ear sensory model, the signal was edited. 23 paper sizes are applicable to China National Standard (CNS) A4 (210 x 297 mm) ------------------ --- Order --------- line I win (please read the notes on the back before filling this page) -ϋ I — ϋ nnn I ^ ϋ ϋ ϋ n ϋ n I · 1 ^ ϋ n ϋ -573294 ^ A7 _B7 V. Description of the invention () (Please read the notes on the back before filling out this page) When calculating the code, first calculate the signal energy of each frequency band and convert the loudness level of the subjective perception of the adult ear. For encoding, it has been proved that it can effectively encode music signals. The natural frequency / (Hz) mapping to Bark scale 6 can be described as: 0 = 681 ^ -1 ^ / = 0 ~ from the middle) (9) For each Bark band, its -3dB bandwidth For a Bark scale, the upper and lower limits, center frequency, and bandwidth of each sensory band can be expressed as follows: fl (n) = 600sinh (^^), (10) 6 / >) = 60〇Sinh (^^), (11) 〇 / c ⑻ = 600sinh (faint), (12) 〇cb \\ {n) = fu {n)-ft {n) (13) It is known from literature research that the speech and audio bandwidth of 4kHz , Can be divided into N (= 1 5) Bark bands, as shown in Table (1). Printed by the Intellectual Property Bureau of the Ministry of Economic Affairs of the Consumer Cooperatives 24. This paper size applies the Chinese National Standard (CNS) A4 specification (210 X 297 mm) 573294 'A7 B7 Table (1): Parameter table of natural frequency mapping to Bark (unit: Hz)
Bark刻度 頻帶下限 頻帶上限 中心頻率 頻寬 i //(0 Λ(0 Λ⑺ cb\v{i) 1 50 152 101 102 2 152 257 204 105 3 257 370 313 113 4 370 493 431 123 5 493 630 560 137 6 630 785 706 155 7 785 961 870 176 8 961 1164 1060 203 9 1164 1400 1278 236 10 1400 1674 1532 274 11 1674 1995 1829 321 12 1995 2372 2177 377 13 2372 2815 2858 443 14 2815 3336 3065 521 15 3336 3950 3631 614 為了將頻譜封包向量轉換成固定維度之 向量，並考慮人耳對不同頻率之聽覺解析度 25 本紙張尺度適用中國國家標準（CNS)A4規格（210 X 297公釐） ------——裝--------訂--------- (請先閱讀背面之注意事項再填寫本頁) 573294 ‘ A7 B7 五、發明說明（） 先 閱 讀 背 面 之 注 意 事 項 再 填 寫 本 頁 不同的特性，本方法將經由維度轉換所得到 的固定Μ個維度之頻譜封包向量，非線性分 割成Ν個頻帶，每一頻帶之頻帶參數及其所 涵蓋之Μ個維度之頻譜封包向量頻率索引值 如表（二）所示。 表(二）：非線性頻率解析度頻帶分割之參數表渾位:Hz)Bark scale band lower limit band upper limit center frequency bandwidth i // (0 Λ (0 Λ⑺ cb \ v (i) 1 50 152 101 102 2 152 257 204 105 3 257 370 313 113 4 370 493 431 123 5 493 630 560 137 6 630 785 706 155 7 785 961 870 176 8 961 1164 1060 203 9 1164 1400 1278 236 10 1400 1674 1532 274 11 1674 1995 1829 321 12 1995 2372 2177 377 13 2372 2815 2858 443 14 2815 3336 3065 521 15 3336 3950 3631 614 In order to convert the spectrum packet vector into a vector of a fixed dimension, and consider the human ear's hearing resolution for different frequencies 25 This paper size applies the Chinese National Standard (CNS) A4 specification (210 X 297 mm) -------- —Installation -------- Order --------- (Please read the precautions on the back before filling this page) 573294 'A7 B7 V. Description of the invention () Read the precautions on the back before Fill in the different characteristics of this page. This method will non-linearly divide the spectrum packet vectors of fixed M dimensions obtained through dimensional conversion into N frequency bands, the frequency band parameters of each frequency band and the spectrum packets of the M dimensions it covers. Vector frequency index The values are shown in Table (2). Table (2): Parameter table for non-linear frequency resolution and frequency band division (Unit: Hz)
Bark刻度 離散頻率指標 中心頻率 頻寬 1 1-1 91 91 2 2-2 182 91 3 3-3 273 91 4 4-4 364 91 5 5-6 500 182 6 7-8 681 182 7 9-10 864 182 8 11-12 1055 182 9 13-15 1273 273 10 16-18 1546 273 11 19-21 1818 273 12 22-25 2136 364 13 26-30 2546 455 14 31-36 3046 546 經濟部智慧財產局員工消費合作社印製 26 本紙張尺度適用中國國家標準（CNS)A4規格（210 X 297公釐） 573294 k A7 B7 五、發明說明（ 15 37-44 3636 727 表（一）與表（二）之差異為，表（一）為連續 性自然頻率到Bark的映對分割；表（二）為以 離散的方式作近似分割。請參閱第四圖（a)及 第四圖（b)，其係本案之離散頻率（表（二））與 傳統連續性自然頻率（表（一））以非線性方式 做頻帶分割後之頻帶與頻寬對照比較圖及本 案之離散頻率（表（二））與傳統連續性自然頻 率（表（一））以非線性方式做頻帶分割後之頻 帶與中心頻率對照比較圖。由圖可知，以表 (二）作非線性分割其頻寬及中心頻率，與式 (9)〜（1 1 )之Bark刻度相近。 經濟部智慧財產局員工消費合作社印製 (二）、非線性分頻頻譜封包之向量量化 (Nonlinear Split-Band SEVQ，NSBSEVQ) 圖（五）所示為本發明之頻譜封包向量量 化之架構圖。固定Μ個維度之頻譜封包向量 r(F) = ，利用預強調濾波器51作等 化處理後，再經由表（二）所述之非線性頻帶 分割（步驟52)，以得致N個維度頻譜封包向 量，接著對每一分割之頻帶所涵蓋之頻率， 求得各頻帶之功率頻譜值（步驟5 3 ): r» = 10*log 丄Σ卜⑺⑺)2*w) Q n ie Band η 27 本紙張尺度適用中國國家標準（CNS)A4規格（210 X 297公釐） -----------^^裝--------訂--------- (請先閱讀背面之注意事項再填寫本頁) \，2 …N (14) 573294 A7 B7 經濟部智慧財產局員工消費合作社印製 五、發明說明（） 其中，％代表其涵蓋之Μ個維度之頻譜 封包向量離散頻率個數，⑷，=1，21}代表 維度為Μ之H (z)功率頻譜。經由上式，我們 將Μ個維度之頻譜封包向量“，以非線性 方式映對成Ν個維度之向量&。因為L P C係 數具備時變（Time Varying)之特性，在（12) 式計算頻帶功率頻譜時，為降低計算複雜 度，忽略了加權濾波器之響應；因為由實驗 得知，在該式中的對數運算（步驟.5 3 )，大幅 降低加權濾波器對碼字搜尋之影響。對原始 編碼之碼字向量{gF)，y = 0，l，...L-l}亦可利用相同之 方式，轉換成N個維度碼字向量 {^7 = 0,1,2,...L-l}(儲存於對數刻度向量碼薄58 中）。利用{〜⑻}(第一目標向量）及 {^^• = 0，1，2，.』-1}(對數刻度向量碼薄58其中之一 碼字向量），分別求得其均值（步驟5 4及使用 碼字向量均值表57)，並將向量分別扣除其 均值，可得到均值為零之向量^(第二目標向 量）及{匕}(基準碼字向量）。為減少編碼搜尋之 計算複雜度，匕}(基準碼字向量）及均值被預 先計算並予以儲存，故在編碼時不需耗費計 算量，對每一音框作碼字搜尋之前，僅需計 算頻譜向量（第二目標向量）即可。而在進行 最佳碼字搜尋時，我們採用利用最小距離準 則，可找到一最佳之碼字％，此一 N個維度 向量之封包形狀與ί〆第二目標向量）最為相 28 本紙張尺度適用中國國家標準（CNS)A4規格（210 X 297公釐） --------------------訂---------線 (請先閱讀背面之注意事項再填寫本頁) 573294 ^ A7 _B7 五、發明說明（） 似（步驟5 6): = = (15) j n=\ 上式中之碼字搜尋需 2 N L 加法及 N L 個乘法運算。經由上述方式搜尋得到最佳之 碼字索引值/之後，該碼字所需之最佳增益 值可取得如下： ⑻ Ά)(16) 八η=1 完成碼字搜尋後，若碼字索引值/代表 最佳碼字，其對應之最佳碼字增益值，在原 始編碼方法採用2位元純量量化（步驟5 5 )。為 對此最佳碼字增益值作純量量化，若G為 原始H VXC編碼器之增益碼薄，ζ代表最佳增 益之索引值如下所示： />argMm{[g^, -2*log10[G0(〇]]2, / = 0-2^1 (17) 完成最佳碼字及增益搜尋後，若以ζ表示 量化過之碼字增益，/=16%'+/:代表最佳碼字索 引值，最後所得到之量化過之殘值頻譜封包 向量(在固定Μ個維度）可表示成： ----------------------訂---------I秦 (請先閱讀背面之注意事項再填寫本頁) 經濟部智慧財產局員工消費合作社印製 三Bark scale discrete frequency index center frequency bandwidth 1 1-1 91 91 2 2-2 182 91 3 3-3 273 91 4 4-4 364 91 5 5-6 500 182 6 7-8 681 182 7 9-10 864 182 8 11-12 1055 182 9 13-15 1273 273 10 16-18 1546 273 11 19-21 1818 273 12 22-25 2136 364 13 26-30 2546 455 14 31-36 3046 546 Intellectual Property Bureau employees of the Ministry of Economic Affairs Printed by the cooperative 26 This paper size applies the Chinese National Standard (CNS) A4 specification (210 X 297 mm) 573294 k A7 B7 V. Description of the invention (15 37-44 3636 727 The difference between Table (1) and Table (2) is Table (1) shows the mapping of continuous natural frequencies to Bark; Table (2) shows the approximate division in a discrete manner. Please refer to Figure 4 (a) and Figure 4 (b), which are the examples in this case. Discrete frequency (table (b)) and traditional continuous natural frequency (table (a)) after a non-linear division of the frequency band and bandwidth comparison chart and the discrete frequency (table (b)) of this case and traditional continuous Natural frequency (Table (1)) Comparison chart of frequency band and center frequency after band division in a non-linear manner . As can be seen from the figure, the bandwidth and center frequency of the non-linear division using Table (2) are similar to the Bark scales of formulas (9) to (1 1). Printed by the Consumer Cooperative of the Intellectual Property Bureau of the Ministry of Economic Affairs (2), Nonlinear Split-Band SEVQ (NSBSEVQ) Figure (5) shows the architecture diagram of the spectral packet vector quantization of the present invention. The spectrum packet vector r (F) with fixed M dimensions =, After using the pre-emphasis filter 51 for equalization processing, the non-linear frequency band segmentation (step 52) described in Table (2) is performed to obtain N-dimensional spectral envelope vectors, and then for each of the divided frequency bands, Frequency, find the power spectrum value of each frequency band (step 5 3): r »= 10 * log 丄 Σ 卜 ⑺⑺) 2 * w) Q n ie Band η 27 This paper size applies Chinese National Standard (CNS) A4 specification ( 210 X 297 mm) ----------- ^^ Installation -------- Order --------- (Please read the precautions on the back before filling this page ) \, 2… N (14) 573294 A7 B7 Printed by the Consumer Cooperatives of the Intellectual Property Bureau of the Ministry of Economic Affairs 5. Description of the invention () Among them,% represents the number of M covered by it The number of discrete frequency spectrum of the vector packet, ⑷, = 1,21} Representative Μ dimension of H (z) power spectrum. Through the above formula, we will map the spectral packet vectors of M dimensions "in a non-linear manner to vectors of N dimensions &. Because the LPC coefficient has the characteristics of time variation, the frequency band is calculated in (12) In the power spectrum, in order to reduce the computational complexity, the response of the weighting filter is ignored; because it is known from experiments that the logarithmic operation in this formula (step .53) greatly reduces the influence of the weighting filter on the codeword search. For the original codeword vector {gF), y = 0, l, ... Ll} can also be converted into N dimensional codeword vectors {^ 7 = 0,1,2, ... in the same way. Ll} (stored in logarithmic scale vector codebook 58). Use {~ ⑻} (first target vector) and {^^ • = 0,1,2,. "-1} (of which logarithmic scale vector codebook 58 is One of the codeword vectors), respectively, to obtain the mean value (step 54 and using the codeword vector mean table 57), and subtracting the average value of the vector, respectively, to obtain a vector with a mean value of zero ^ (second target vector) and { (Reference codeword vector). In order to reduce the computational complexity of the code search, (reference codeword vector) and the mean are Calculate and store it, so there is no need to calculate when coding, just calculate the spectral vector (the second target vector) before searching for the codeword of each sound frame. When searching for the best codeword, We use the minimum distance criterion to find an optimal codeword%. The packet shape of this N-dimensional vector is most similar to the second target vector. 28 This paper scale applies the Chinese National Standard (CNS) A4 specification ( 210 X 297 mm) -------------------- Order --------- line (Please read the precautions on the back before filling this page) 573294 ^ A7 _B7 V. Description of the invention () (Step 5 6): = = (15) jn = \ The codeword search in the above formula requires 2 NL addition and NL multiplication operations. The best search is obtained through the above method. After the codeword index value /, the optimal gain value required by the codeword can be obtained as follows: 16 Ά) (16) Eight η = 1 After the codeword search is completed, if the codeword index value / represents the best codeword, its Corresponding to the optimal codeword gain value, 2-bit scalar quantization is used in the original encoding method (step 5 5). To make this optimal codeword gain value scalar Quantization, if G is the gain codebook of the original H VXC encoder, ζ represents the index value of the best gain as follows: / > argMm {[g ^, -2 * log10 [G0 (〇]] 2, / = 0-2 ^ 1 (17) After searching for the best codeword and gain, if quantized codeword gain is represented by ζ, / = 16% '+ /: represents the best codeword index value, and the resulting quantization The passing residual value spectrum packet vector (in a fixed M dimension) can be expressed as: ---------------------- Order --------- I Qin (Please read the notes on the back before filling out this page) Printed by the Consumer Cooperatives of the Intellectual Property Bureau of the Ministry of Economic Affairs
61 + k 氺 0 GO( II 估化 平量立曰 古口 質量語 σα向文 CD :估中 評段 -v=口 士口了四 、為以 aj 我 *\ar 和 質 品 析度 分雜 度複 雜算 算及 聲 女 段 兩 聲 男 段 兩 9 2 本紙張尺度適用中國國家標準（CNS)A4規格（210 X 297公釐） 573294 1 A7 B7 五、發明說明（ 作測試樣本，以客觀形式評估量化誤差，我 們採用被廣為使用在頻率域之誤差評估方 式：“頻譜失真度 ” （Spectral Distance，SD) 的量測方式，在可變維度P或固定維度Μ， 對量化前後之聽覺加權頻譜封包向量作均方 誤差計算。因為MPEG - 4 HVXC解碼端，將固 定維度向量碼字轉成可變維度之振幅頻譜， 以此頻譜去合成語音信號；且在編碼端’若 語音編碼位元率為4 k b p s時，需對基本層之 諧波頻譜封包向量量化誤差進一步作增強層 (Enhancement Layer) ^，亦在時變維度上 執行，綜合以上兩個因素，我們採用在時變 維度P作量化品質評估。若在時變維度P作 評估，對某一音框信號，其頻譜失真度切⑷如 下所示： 1/2 19) 1尸 -Σ {2° 1〇§10 K〇^(〇] - 20 log10 [r(0 w(0]}2 Γ 對長度為K音框之語音信號，最終之客 觀品質評估頻譜失真度，為對所有音框之頻 譜失真值作平均求得，如下： 經濟部智慧財產局員工消費合作社印製 (請先閱讀背面之注意事項再填寫本頁) 切=丄( 2 〇 >61 + k 氺 0 GO (II Estimation of averaging Ligu Gulangyu σα Xiangwen CD: In the evaluation, the paragraph -v = 口 口 口 了 4. In order to use aj I * \ ar and quality resolution Complex calculations and two voices for male and female sections 9 2 This paper size is applicable to Chinese National Standard (CNS) A4 (210 X 297 mm) 573294 1 A7 B7 V. Description of the invention (for test samples and evaluation in objective form For the quantization error, we adopt a widely used error evaluation method in the frequency domain: the "Spectral Distance" (SD) measurement method. In the variable dimension P or the fixed dimension M, the auditory weighted spectrum before and after quantization is weighted. The packet vector is used to calculate the mean square error. Because the MPEG-4 HVXC decoder converts a fixed-dimensional vector codeword into a variable-dimensional amplitude spectrum and uses this spectrum to synthesize a speech signal; and at the encoding end, if the speech encoding bit rate When it is 4 kbps, the harmonic quantization error of the harmonic spectrum packet vector of the base layer needs to be further enhanced layer (Enhancement Layer) ^. It is also implemented in the time-varying dimension. Based on the above two factors, we use the P is used to evaluate the quality of quantification. If evaluated in the time-varying dimension P, the spectral distortion of a sound frame signal is cut as follows: 1/2 19) 1 corpus-Σ {2 ° 1〇§10 K〇 ^ (〇]-20 log10 [r (0 w (0)} 2 Γ For speech signals with a length of K frames, the final objective quality is to evaluate the spectral distortion degree, which is obtained by averaging the spectral distortion values of all the frames. , As follows: Printed by the Consumer Cooperatives of the Intellectual Property Bureau of the Ministry of Economic Affairs (please read the precautions on the back before filling out this page) Cut = 丄 (2 〇 >
Ns k=\ 請參閱第六圖（a)及第六圖（b)，其係一 音框語音信號（基週值為5 4、對應之諧波個 數P為2 7 )之信號波形圖及頻譜圖，而第六 30 本紙張尺度適用中國國家標準（CNS)A4規格（210 X 297公釐） 573294 · A7 B7 五、發明說明（ 請 先 閱 讀 背 Φ 之 注 項Ns k = \ Please refer to Figure 6 (a) and Figure 6 (b), which are the signal waveforms of a sound box voice signal (base cycle value is 5 4 and the corresponding harmonic number P is 2 7) And spectrogram, and the 30th paper size is applicable to China National Standard (CNS) A4 (210 X 297 mm) 573294 · A7 B7 V. Description of the invention (please read the note on the back Φ first
f I I I I I 圖（c )係為以本案之方法與以習用Η V X C向量 量化法對該音框語音信號進行量化後之頻譜 封包比較圖。第七圖（a )及第七圖（b )係為一 段語音信號之信號波形圖及基週變化曲線 圖，而第七圖（c )係為以本案之方法與以習用 Η V X C向量量化法對該段語音信號進行量化 後之頻譜失真比較圖。表（三）表示對此四段 中文語音樣本，作客觀品質評估所得到之實 驗結果。由以上圖表顯示，本案之方法，以 上述S D頻譜失真測量方式評估，具備較小 之量化誤差。 表(三）：不同的頻譜封包向量量化之SD量化誤差比較表(單位:dB) 訂 量化模式 男聲#1 男聲#2 女聲#1 女1 1 m ! 1 HVXC 3.87 3.46 4.11 3.： 1 〇 1 k 本案 3.50 3.26 3.72 2. .8 # -1_ 經 濟 部 智 慧 財 產 局 員 工 消 費 合 作 社 印 製 對原始量化方式及本案所提出之量化方 法，對一個音框語音信號頻譜封包量化，作 計算複雜度理論分析比較，結果如表（四）所 示。由表中數據可知，本案之量化方式大幅 降低計算複雜度約9倍。 31 本紙張尺度適用中國國家標準（CNS)A4規格（210 X 297公釐） 573294 A7 B7 經濟部智慧財產局員工消費合作社印製 五、發明說明（） 表(四）：不同的頻譜封包向量量化之計算複雜度比較表 量化模式 加減法運算（A/D) 乘除法運算(M/D) 實例* 加減法數 乘除法數 HVXC 2ML (3M+2)L+M 22528 34348 本案 2NL+2N+1 NL+2 7711 3842 *以从=44，#=15，乙=256爲例 綜上所述，本案利用人類耳朵之感官特 性，對聲音信號振幅之感知具備非線性刻度 及不同頻率具備不同頻率解析度之特性，將 各音框信號經過線性預估所得之殘值信號的 頻譜封包，以非線性的方式分割成固定個數 之頻帶，並計算各感官頻帶内所含諧波頻譜 之功率分貝值後，再作向量量化，大幅降低 計算量及量化誤差，且所得之編碼參數與現 行標準規範之解碼器相容，不需要改變解碼 系統之架構，有利於編碼系統之軟硬體實 現，有效改善習知技術之缺失，是故具有產 業價值，進而達成發展本案之目的。 本案得由熟悉本技藝之人士任施匠思而 為諸般修飾，然皆不脫如附申請專利範圍所 32 本紙張尺度適用中國國家標準（CNS)A4規格（210 X 297公釐） --------------------訂---------線 (請先閱讀背面之注意事項再填寫本頁) 573294 ^ A7 _B7 五、發明說明（） 者 護 保 欲 (請先閱讀背面之注意事項再填寫本頁) 經濟部智慧財產局員工消費合作社印製 3 3 本紙張尺度適用中國國家標準（CNS)A4規格（210 X 297公釐）f I I I I I (c) is a comparison diagram of the frequency spectrum packet after quantizing the voice signal of the sound box by the method of this case and the conventional Η V X C vector quantization method. The seventh diagram (a) and the seventh diagram (b) are a signal waveform diagram and a base cycle change curve of a speech signal, and the seventh diagram (c) is the method of the present case and the conventional Η VXC vector quantization method Comparison chart of spectral distortion after quantizing the speech signal. Table (3) shows the experimental results obtained by objectively evaluating the four Chinese speech samples. The above chart shows that the method in this case is evaluated with the above SD spectral distortion measurement method and has a smaller quantization error. Table (3): Comparison table of SD quantization errors for vector quantization of different spectrum packets (Unit: dB) Custom quantization mode Male voice # 1 Male voice # 2 Female voice # 1 Female 1 1 m! 1 HVXC 3.87 3.46 4.11 3 .: 1 〇1 k This case 3.50 3.26 3.72 2. .8 # -1_ The consumer cooperative of the Intellectual Property Bureau of the Ministry of Economic Affairs printed the original quantization method and the quantization method proposed in this case, and quantified the spectral envelope of a voice frame voice signal for theoretical analysis of computational complexity. The comparison is shown in Table (4). According to the data in the table, the quantification method in this case significantly reduced the calculation complexity by about 9 times. 31 This paper size applies the Chinese National Standard (CNS) A4 specification (210 X 297 mm) 573294 A7 B7 Printed by the Consumer Cooperatives of the Intellectual Property Bureau of the Ministry of Economic Affairs 5. Description of the invention () Table (4): Vectorization of different spectrum envelopes Computational complexity comparison table Quantization mode Addition and subtraction (A / D) Multiplication and division (M / D) Example * Addition and subtraction number Multiplication and division number HVXC 2ML (3M + 2) L + M 22528 34348 This case 2NL + 2N + 1 NL + 2 7711 3842 * Taking from = 44, # = 15, B = 256 as an example In summary, this case uses the sensory characteristics of the human ear to have a non-linear scale for the perception of the amplitude of sound signals and different frequencies for different frequencies. Characteristics, the spectral envelope of the residual signal obtained by linear estimation of each sound frame signal is divided into a fixed number of frequency bands in a non-linear manner, and the power decibel value of the harmonic spectrum contained in each sensory frequency band is calculated Then, vector quantization is performed to greatly reduce the calculation amount and quantization error, and the obtained encoding parameters are compatible with the decoders of the current standard specifications. It is not necessary to change the structure of the decoding system, which is beneficial to the hardware and software of the encoding system It is realized that the lack of effective knowledge technology is of industrial value and thus achieves the purpose of developing this case. This case may be modified by any person who is familiar with this technology, but it is not different from the scope of the patent application. 32 The paper size is applicable to the Chinese National Standard (CNS) A4 (210 X 297 mm) --- ----------------- Order --------- line (please read the notes on the back before filling this page) 573294 ^ A7 _B7 V. Description of the invention ( ) For protection (please read the notes on the back before filling out this page) Printed by the Consumer Cooperatives of the Intellectual Property Bureau of the Ministry of Economic Affairs 3 3 This paper size applies to China National Standard (CNS) A4 (210 X 297 mm)