开通博客赚积分
发布资源赚积分
分类

源码开发语言/平台
当前位置：首页 > 源码/资料 > 多媒体 > 语音合成与识别 > 查看源码
FE_pitch.cpp
资源名称：speech_analysis.rar　[点击查看]
上传用户：italyroyal
上传日期：2013-05-06
资源大小：473k
文件大小：14k
源码类别：

语音合成与识别

开发平台：

Visual C++

FE_pitch.cpp：源码内容
  1. ///////////////////////////////////////////////////////////////////////////////
  2. // This is a part of the Feature program.
  3. // Version: 1.0
  4. // Date: February 22, 2003
  5. // Programmer: Oh-Wook Kwon
  6. // Copyright(c) 2003 Oh-Wook Kwon. All rights reserved. owkwon@ucsd.edu
  7. ///////////////////////////////////////////////////////////////////////////////
  9. #include "StdAfx.h"
  10. #include "FE_feature.h"
  13. struct FePitchSeg /* Pitch detection parameters */
  14. {
  15. EVusType v;
  16. float f;
  17. };
  20. int Fe::pitch_basic(short *sample, int sampleN, int sampleRate, int shiftSize, vector<EVusType>& vusA, vector<float>& pitchA)
  21. {
  22. int i,n;
  23. int maxFrameSize = 2*sampleRate/MIN_PITCH_FREQ; /* 40 ms */
  24. int minFrameSize = 3*sampleRate/MAX_PITCH_FREQ; /* 12 ms */
  25. int num_frames = sampleN/shiftSize;
  26. int lpfLen = (sampleRate/(2*MAX_PITCH_FREQ)); if(lpfLen%2 == 0) lpfLen += 1;
  27. float maxDeltaPitch = (MAX_PITCH_CHANGE*MAX_PITCH_FREQ); /* 50 Hz for 10 ms */
  28. #ifdef _DEBUG
  29. #define TEST_FRAME_LEN 300
  30. EVusType vusTmpA[TEST_FRAME_LEN];
  31. int k;
  32. #endif
  34. pitchA.resize(num_frames+1);
  35. for(i=0;i<=num_frames;i++) pitchA[i]=0;
  36. if(!(sampleN > 0 && sampleRate > 0)) return false;
  37. if(sampleN<maxFrameSize) return false;
  39. vector<float> amdfA(maxFrameSize);
  40. vector<float> frameA(maxFrameSize);
  41. vector<float> p(num_frames+1+PITCH_BUF_SIZE/2);
  42. FePitchSeg bufA[PITCH_BUF_SIZE];
  43. FePitchSeg curr;
  45. if(vusA.size()==0) vus_basic(sample, sampleN, GetFrameSize(), vusA);
  46. pitch_low_pass_filter(sample, sampleN, lpfLen);
  47. for(i=0;i<PITCH_BUF_SIZE;i++) bufA[i].v = FRM_SILENCE;
  49. int nCurrPos = (PITCH_BUF_SIZE-1)/2;
  50. int PITCH_BUF_SIZE2 = (PITCH_BUF_SIZE-1)/2;
  51. m_meanPitch = 0;
  52. m_pitchFrameN = 0;
  53. for(n=0;n < PITCH_BUF_SIZE/2; n++){
  54. curr.v=vusA[n];
  55. if(curr.v == FRM_VOICED){
  56. for(i=0;i<maxFrameSize;i++) frameA[i]=sample[n*shiftSize+i];
  57. m_window.Windowing(&frameA[0],maxFrameSize,WIN_HANNING);
  58. curr.f = pitch_find(&frameA[0], maxFrameSize, &p[0], n, &amdfA[0], (int)(shiftSize/(float)sampleRate), sampleRate);
  59. }
  60. else{
  61. curr.f = 0;
  62. }
  63. p[n] = curr.f;
  65. for(i=0;i<PITCH_BUF_SIZE-1;i++){
  66. bufA[i] = bufA[i+1];
  67. }
  68. bufA[PITCH_BUF_SIZE-1] = curr;
  69. }
  71. int frameN=0;
  72. int blockSize = maxFrameSize;
  73. for(;n*shiftSize+blockSize < sampleN; n++){
  74. if(!CheckWinMessage()) break;
  75. ShowProgress((int)((n*shiftSize*100.0)/sampleN));
  77. curr.v=vusA[n];
  78. if(curr.v == FRM_VOICED){
  79. for(i=0;i<blockSize;i++) frameA[i]=sample[n*shiftSize+i];
  80. m_window.Windowing(&frameA[0],blockSize,WIN_HANNING);
  81. curr.f = pitch_find(&frameA[0], blockSize, &p[0], n, &amdfA[0], (int)(shiftSize/(float)sampleRate), sampleRate);
  82. if(curr.f == 0){
  83. //printf("Correcting pitch to zeron");
  84. curr.v = FRM_UNVOICED;
  85. }
  86. }
  87. else{
  88. curr.f = 0;
  89. }
  91. p[n] = curr.f;
  93. for(i=0;i<PITCH_BUF_SIZE-1;i++){
  94. bufA[i] = bufA[i+1];
  95. }
  96. bufA[PITCH_BUF_SIZE-1] = curr;
  98. int sum;
  99. for(i=1,sum=0;i<PITCH_BUF_SIZE-1;i++) sum+=((bufA[i].v == FRM_VOICED)?1:0);
  100. if(sum>=PITCH_BUF_SIZE-3 && (bufA[nCurrPos].v == FRM_UNVOICED || bufA[nCurrPos].v == FRM_SILENCE)) {
  101. if(bufA[nCurrPos].v == FRM_UNVOICED){
  102. //printf("Correcting FRM_UNVOICED->FRM_VOICEDn");
  103. }
  104. else{
  105. //printf("Correcting FRM_SILENCE->FRM_VOICEDn");
  106. }
  107. bufA[nCurrPos].v = FRM_VOICED;
  108. bufA[nCurrPos].f = (bufA[nCurrPos-1].f+bufA[nCurrPos+1].f)/2;
  109. p[n-(PITCH_BUF_SIZE-1-nCurrPos)] = bufA[nCurrPos].f;
  110. }
  112. for(i=0,sum=0;i<PITCH_BUF_SIZE;i++) sum+=((bufA[i].v == FRM_VOICED)?1:0);
  113. if(sum==PITCH_BUF_SIZE){
  114. int peak[PITCH_BUF_SIZE];
  115. int lmin = -1;
  116. int lmax = PITCH_BUF_SIZE;
  117. for(i=1;i<PITCH_BUF_SIZE;i++){
  118. float delta = (float)fabs(bufA[i].f - bufA[i-1].f);
  119. if(delta>maxDeltaPitch){
  120. peak[i] = 1;
  121. if(lmin < 0) lmin = i;
  122. lmax = i;
  123. }
  124. else{
  125. peak[i] = 0;
  126. }
  127. }
  129. if(lmax-lmin <= 2 && lmin >= 3 && lmax < PITCH_BUF_SIZE){
  130. //printf("Correcting pitchn");
  131. for(i=lmin; i<lmax;i++){
  132. bufA[i].f = (bufA[lmin-1].f+bufA[lmax].f)/2;
  133. p[n-(PITCH_BUF_SIZE-1-i)] = bufA[i].f;
  134. }
  135. }
  136. }
  138. for(i=1,sum=0;i<PITCH_BUF_SIZE-1;i++) sum+=((bufA[i].v == FRM_UNVOICED || bufA[i].v == FRM_SILENCE)?1:0);
  139. if(sum>=PITCH_BUF_SIZE-3 && bufA[nCurrPos].v == FRM_VOICED) {
  140. //printf("Correcting FRM_VOICED->FRM_UNVOICEDn");
  141. bufA[nCurrPos].v = FRM_UNVOICED;
  142. bufA[nCurrPos].f = 0;
  143. p[n-(PITCH_BUF_SIZE-1-nCurrPos)] = bufA[nCurrPos].f;
  145. //printf("Correcting Pitch historyn");
  146. for(i=nCurrPos+1;i<PITCH_BUF_SIZE;i++){
  147. curr.v=vusA[n];
  148. if(curr.v == FRM_VOICED){
  149. int k;
  150. for(k=0;k<blockSize;k++) frameA[k]=sample[n*shiftSize+(i-3)*shiftSize+k];
  151. m_window.Windowing(&frameA[0],blockSize,WIN_HANNING);
  152. curr.f = pitch_find(&frameA[0], blockSize, &p[0], n-(PITCH_BUF_SIZE-1-i), &amdfA[0], (int)(shiftSize/(float)sampleRate), sampleRate);
  153. if(curr.f == 0){
  154. //printf("Correcting pitch to zeron");
  155. curr.v = FRM_UNVOICED;
  156. }
  157. }
  158. else{
  159. curr.f = 0;
  160. }
  161. p[n-(PITCH_BUF_SIZE-1-i)] = curr.f;
  162. bufA[i] = curr;
  163. }
  164. }
  166. if(n < num_frames){
  167. pitchA[frameN] = bufA[nCurrPos].f;
  168. frameN++;
  169. }
  171. /* use variable block size */
  172. if(bufA[nCurrPos].f > 0)
  173. blockSize = my_max(minFrameSize, my_min(maxFrameSize,(int)(3*sampleRate/(bufA[nCurrPos].f/3)) ));
  174. else
  175. blockSize = maxFrameSize;
  176. }
  178. #ifdef _DEBUG
  179. for(k=0;k<my_min(TEST_FRAME_LEN,vusA.size());k++){
  180. vusTmpA[k]=vusA[k];
  181. }
  182. #endif
  183. vector<EVusType> vusOrgA(num_frames+1);
  184. for(i=0;i<=num_frames;i++) vusOrgA[i] = FRM_SILENCE;
  185. frameN = 0;
  186. for(n=0;n*shiftSize+blockSize < sampleN; n++){
  187. EVusType v=vusA[n];
  188. if(n < num_frames){
  189. if(v == FRM_VOICED && pitchA[n]==0)
  190. vusOrgA[frameN] = FRM_UNVOICED;
  191. else
  192. vusOrgA[frameN] = v;
  193. frameN++;
  194. }
  195. vusA[n] = vusOrgA[n];
  196. }
  197. #ifdef _DEBUG
  198. for(k=0;k<my_min(TEST_FRAME_LEN,vusA.size());k++){
  199. vusTmpA[k]=vusA[k];
  200. }
  201. #endif
  203. /* correct isolated pitch */
  204. for(i=0;i<frameN;i++){
  205. if(vusA[i]==FRM_VOICED && pitchA[i]==0) vusA[i]=FRM_UNVOICED;
  206. }
  207. vus_remove_short_segments(vusA); /* remove short segments */
  208. for(i=0;i<frameN;i++){
  209. if(vusA[i]==FRM_UNVOICED || vusA[i]==FRM_SILENCE) pitchA[i]=0;
  210. }
  212. /* we have to shift a few frames because a longer block size is used for pitch estimation */
  213. int feFrameSize=GetFrameSize();
  214. int corrFrameN=(int)(0.5*(maxFrameSize-feFrameSize)/(float)feFrameSize);
  215. for(i=frameN-1;i>=corrFrameN;i--){
  216. pitchA[i]=pitchA[i-corrFrameN];
  217. }
  219. /* spike removal */
  220. pitch_remove_spike(pitchA);
  222. /* low-pass filter, x=filter([1 2 1]/4, [1], x); */
  223. pitch_linear_filter(pitchA);
  225. return frameN;
  226. }
  229. bool Fe::pitch_amdf(float *sample, float *amdfA, int blockSize, int pmin, int pmax)
  230. {
  231. int s,n;
  232. for(s=pmin;s<=pmax;s++){
  233. amdfA[s] = 0;
  234. for(n=0;n<blockSize;n++){
  235. if(n+s < blockSize){
  236. amdfA[s] += my_abs(sample[n]-sample[n+s]);
  237. }
  238. else{
  239. amdfA[s] += my_abs(sample[n]);
  240. }
  241. }
  242. }
  243. return true;
  244. }
  247. float Fe::pitch_find(float *sample, int blockSize, float *pitchA, int t, float *amdfA, int shiftSize, int sampleRate)
  248. {
  249. int pmin = sampleRate/MAX_PITCH_FREQ;
  250. int pmax = sampleRate/MIN_PITCH_FREQ;
  251. int i;
  253. float lastPitch = (float)0;
  254. int lastTime = 0;
  255. for(i=0;i<t;i++){
  256. if(pitchA[i]>0) {
  257. lastPitch = pitchA[i];
  258. lastTime = i;
  259. }
  260. }
  261. int startSegTime=t;
  262. for(i=t-1;i>=0;i--){
  263. if(pitchA[i]==0) {
  264. startSegTime=i+1;
  265. break;
  266. }
  267. }
  269. if(t>=167 && t<180){
  270. int aaa=1;
  271. aaa++;
  272. }
  273. float lowF = (float)MIN_PITCH_FREQ;
  274. float highF = (float)MAX_PITCH_FREQ;
  275. float maxDeltaPitch = (MAX_PITCH_CHANGE*MAX_PITCH_FREQ); // 40Hz for 10ms
  276. if(lastPitch > 0){
  277. if(t >= 3 && lastTime == t-1 && (pitchA[t-2] == (float)0 || pitchA[t-3] == (float)0)){
  278. // Start of voice
  279. maxDeltaPitch = (2*MAX_PITCH_CHANGE*lastPitch);
  280. }
  281. else{
  282. // Middle of voice
  283. maxDeltaPitch = (MAX_PITCH_CHANGE*lastPitch);
  284. }
  285. }
  287. float deltaPitch = maxDeltaPitch;
  288. if(lastPitch > 0){
  289. int dur = (t-lastTime);
  290. deltaPitch = dur*maxDeltaPitch;
  291. lowF = my_max(MIN_PITCH_FREQ,lastPitch - deltaPitch);
  292. highF = my_min(MAX_PITCH_FREQ,lastPitch + deltaPitch);
  293. pmin = (int)(sampleRate/highF);
  294. pmax = (int)(sampleRate/lowF);
  295. }
  296. int smin;
  297. int smax;
  299. pitch_amdf(sample, amdfA, blockSize, pmin, pmax);
  300. pitch_find_minmax(amdfA, pmin, pmax, &smin, &smax);
  302. float avgAmdf=0;
  303. for(i=pmin;i<=pmax;i++) avgAmdf += amdfA[i];
  304. avgAmdf /= (pmax-pmin+1);
  305. float amdf=amdfA[smin];
  307. if(t>=167 && t<180){
  308. int aaa=1;
  309. aaa++;
  310. }
  311. float currPitch;
  312. bool reliable=false;
  313. if(pmin < smin && smin < pmax && amdf<avgAmdf*F0_SHARP_TH_1){
  314. currPitch = 1/(float)smin * sampleRate;
  315. if(amdf<avgAmdf*0.9) reliable=true;
  316. }
  317. else{
  318. // try once more with increased range
  319. if(smin == pmax){
  320. lowF = my_max(MIN_PITCH_FREQ,lowF-deltaPitch);
  321. }
  322. if(smin == pmin){
  323. highF = my_min(MAX_PITCH_FREQ,highF+deltaPitch);
  324. }
  325. pmin = (int)(sampleRate/highF);
  326. pmax = (int)(sampleRate/lowF);
  328. pitch_amdf(sample, amdfA, blockSize, pmin, pmax);
  329. pitch_find_minmax(amdfA, pmin, pmax, &smin, &smax);
  331. avgAmdf=0;
  332. for(i=pmin;i<=pmax;i++) avgAmdf += amdfA[i];
  333. avgAmdf /= (pmax-pmin+1);
  334. amdf=amdfA[smin];
  336. if((pmin < smin && smin < pmax) && amdf<avgAmdf*F0_SHARP_TH_2){
  337. currPitch = 1/(float)smin * sampleRate;
  338. }
  339. else{
  340. currPitch = (float)0;
  341. }
  342. if(amdf<avgAmdf*0.5) reliable=true;
  343. }
  345. int gender=0;
  346. if(m_pitchFrameN>20){
  347. if(m_meanPitch>200)
  348. gender=1;
  349. else if(m_meanPitch<150)
  350. gender=-1;
  351. }
  352. /* Correct F0 doubling/halving error */
  353. float orgMag=pitch_fft_one_freq(sample,blockSize,smin);
  354. bool checkDoubling = ((currPitch>F0_DOUBLE_FREQ_TH) || (currPitch>1.6*m_meanPitch));
  355. bool checkHalving = ((currPitch>0) && (currPitch<F0_HALVE_FREQ_TH || currPitch<0.6*m_meanPitch));
  356. bool checkGender = ((gender==1 && currPitch<150) || (gender== -1 && currPitch>200));
  357. float f0_double_th = (float)F0_DOUBLE_TH;
  358. float f0_halve_th = (float)F0_HALVE_TH;
  359. if(checkGender != 0){
  360. if(checkDoubling){
  361. f0_double_th *= 1;
  362. f0_halve_th *= (float)0.2;
  363. }
  364. else if(checkHalving){
  365. f0_double_th *= (float)0.5;
  366. f0_halve_th *= 1;
  367. }
  368. }
  369. if(t>=252){
  370. int aaa=1;
  371. aaa++;
  372. }
  374. /* Apply correction when the estimated F0 is very different at the start of segment */
  375. /* and magnitude of the frequency is large enough  */
  376. if(t-startSegTime <= 5 && orgMag > 0.5*blockSize && (checkDoubling || checkHalving || checkGender)){ 
  377. float hypoDouble, hypoHalve;
  378. hypoDouble=pitch_fft_one_freq(sample,blockSize,smin/2);
  379. hypoHalve=pitch_fft_one_freq(sample,blockSize,smin*2);
  380. if(checkGender==0 && orgMag>hypoDouble*F0_NO_DOUBLE_TH && orgMag>hypoHalve*F0_NO_HALVE_TH){
  381. ;
  382. }
  383. else if(checkDoubling){
  384. if(orgMag > f0_double_th*hypoDouble && hypoHalve > f0_halve_th*orgMag)
  385. currPitch /= 2;
  386. }
  387. else if(checkHalving){
  388. if(orgMag < f0_double_th*hypoDouble)
  389. currPitch *= 2;
  390. }
  391. }
  393. /* update the mean pitch frequency */
  394. if(reliable){
  395. float lambda;
  396. if(m_pitchFrameN<10){
  397. lambda = 1-1/(float)(m_pitchFrameN+1);
  398. }
  399. else{
  400. lambda = (float)lambdaPitchTrack;
  401. }
  402. m_meanPitch = lambda*m_meanPitch + (1-lambda)*currPitch;
  403. m_pitchFrameN++;
  404. }
  405. return currPitch;
  406. }
  409. float Fe::pitch_fft_one_freq(float *sample, int blockSize, int p)
  410. {
  411. int i;
  412. float xre=0, xim=0;
  413. for(i=0;i<blockSize;i++){
  414. xre += sample[i]*cos(2*M_PI*i/(float)p);
  415. xim += sample[i]*sin(2*M_PI*i/(float)p);
  416. }
  417. return (float)sqrt(xre*xre + xim*xim);
  418. }
  421. bool Fe::pitch_find_minmax(float *amdfA, int pmin, int pmax, int* smin, int* smax)
  422. {
  423. int s;
  424. *smin = pmin;
  425. *smax = pmin;
  426. float amin = amdfA[pmin];
  427. float amax = amdfA[pmin];
  428. for(s=pmin;s<=pmax;s++){
  429. if(amdfA[s] < amin){
  430. amin = amdfA[s];
  431. *smin = s;
  432. }
  433. else if(amdfA[s] > amax){
  434. amax = amdfA[s];
  435. *smax = s;
  436. }
  437. }
  438. return true;
  439. }
  442. bool Fe::pitch_low_pass_filter(short *sample, int sampleN, int lpfLen)
  443. {
  444. int i,n,sum,kk;
  445. vector<int> tmp(sampleN);
  446. for(n=0;n<lpfLen/2;n++){
  447. sum=0;
  448. kk=0;
  449. for(i=-lpfLen/2;i<=lpfLen/2;i++){
  450. sum += (((n+i) >= 0) ? sample[n+i] : 0);
  451. kk += (((n+i) >= 0) ? 1 : 0);
  452. }
  453. tmp[n] = sum/kk;
  454. }
  456. for(n=lpfLen/2;n<sampleN-lpfLen/2;n++){
  457. sum = 0;
  458. for(i=-lpfLen/2;i<=lpfLen/2;i++) sum += sample[n+i];
  459. tmp[n] = sum/lpfLen;
  460. }
  462. for(n=sampleN-lpfLen/2;n<sampleN;n++){
  463. sum=0;
  464. kk=0;
  465. for(i=-lpfLen/2;i<=lpfLen/2;i++){
  466. sum += (((n+i) < sampleN) ? sample[n+i] : 0);
  467. kk += (((n+i) >= 0) ? 1 : 0);
  468. }
  469. tmp[n] = sum/kk;
  470. }
  472. /* clipping to 16 bit integer */
  473. for(n=0;n<sampleN;n++){
  474. sample[n] = my_min(my_max(SHRT_MIN, tmp[n]), SHRT_MAX);
  475. }
  477. return true;
  478. }
  481. int Fe::pitch_remove_spike(vector<float>& pitchA)
  482. {
  483. int i;
  484. float maxChange=(float)MAX_PITCH_CHANGE_POSTPROC;
  485. int frameN=pitchA.size();
  486. for(i=2;i<frameN-2;i++){
  487. if(pitchA[i-1]==0 && pitchA[i+1]==0){
  488. pitchA[i]=0;
  489. }
  490. else if(pitchA[i-1]>0 && pitchA[i+1]>0){
  491. float ftmp=GetMedian(&pitchA[i-2],5);
  492. if(fabs(ftmp-pitchA[i])>maxChange*pitchA[i]){
  493. pitchA[i]=(pitchA[i-1]+pitchA[i+1])/2;
  494. }
  495. }
  496. else if(i>=5 && pitchA[i]>0 && pitchA[i-1]>0 && pitchA[i-2]>0 && pitchA[i+1]==0){
  497. float ftmp=GetMedian(&pitchA[i-5],5);
  498. if(fabs(ftmp-pitchA[i])>maxChange*pitchA[i]){
  499. pitchA[i]=my_max(-pitchA[i-2]+my_min(2*pitchA[i-1],MAX_PITCH_FREQ), 0);
  500. }
  501. }
  502. else if(i<frameN-5 && pitchA[i]>0 && pitchA[i+1]>0 && pitchA[i+2]>0 && pitchA[i-1]==0){
  503. float ftmp=GetMedian(&pitchA[i],5);
  504. if(fabs(ftmp-pitchA[i])>maxChange*pitchA[i]){
  505. pitchA[i]=my_max(my_min(2*pitchA[i+1],MAX_PITCH_FREQ)-pitchA[i+2], 0);
  506. }
  507. }
  508. }
  509. return frameN;
  510. }
  513. int Fe::pitch_linear_filter(vector<float>& pitchA)
  514. {
  515. int i;
  516. int frameN=pitchA.size();
  517. vector<float> x;
  518. x.resize(frameN);
  519. for(i=0;i<frameN;i++) x[i]=pitchA[i];
  521. for(i=2;i<frameN-2;i++){
  522. if(x[i-2]>0 && x[i-1]>0 && x[i+1]>0 && x[i+2]>0){
  523. pitchA[i]=(x[i-1]+2*x[i]+x[i+1])/4;
  524. }
  525. }
  526. return frameN;
  527. }