realtime-1stpass.c

julius version 4.12.about sound recognition.

    } else {
      init_param(mfcc);
    }
    /* フレ〖ムごとのパラメ〖タベクトル瘦赂の挝拌を澄瘦 */
    /* あとで涩妥に炳じて凯墓される */
    if (param_alloc(mfcc->param, 1, mfcc->param->veclen) == FALSE) {
      j_internal_error("ERROR: segmented: failed to allocate memory for rest param\n");
    }
    /* フレ〖ム眶をリセット */
    /* reset frame count */
    mfcc->f = 0;
  }
  /* 洁洒した param 菇陇挛のデ〖タのパラメ〖タ房を不读モデルとチェックする */
  /* check type coherence between param and hmminfo here */
  if (recog->jconf->input.paramtype_check_flag) {
    for(am=recog->amlist;am;am=am->next) {
      if (!check_param_coherence(am->hmminfo, am->mfcc->param)) {
	jlog("ERROR: input parameter type does not match AM\n");
	return FALSE;
      }
    }
  }

  /* 纷换脱のワ〖クエリアを洁洒 */
  /* prepare work area for calculation */
  if (recog->jconf->input.type == INPUT_WAVEFORM) {
    reset_mfcc(recog);
  }
  /* 不读锑刨纷换脱キャッシュを洁洒 */
  /* prepare cache area for acoustic computation of HMM states and mixtures */
  for(am=recog->amlist;am;am=am->next) {
    outprob_prepare(&(am->hmmwrk), r->maxframelen);
  }

#ifdef BACKEND_VAD
  if (recog->jconf->decodeopt.segment) {
    /* initialize segmentation parameters */
    spsegment_init(recog);
  }
#else
  recog->triggered = FALSE;
#endif

#ifdef DEBUG_VTLN_ALPHA_TEST
  /* store speech */
  recog->speechlen = 0;
#endif

  return TRUE;
}

/** 
 * <JA>
 * @brief  不兰侨妨からパラメ〖タベクトルを纷换する.
 * 
 * 岭帽疤で艰り叫された不兰侨妨からMFCCベクトルを纷换する.
 * 纷换冯蔡は mfcc->tmpmfcc に瘦赂される. 
 * 
 * @param mfcc [i/o] MFCC纷换インスタンス
 * @param window [in] 岭帽疤で艰り叫された不兰侨妨デ〖タ
 * @param windowlen [in] @a window の墓さ
 * 
 * @return 纷换喇根箕·TRUE を手す. デルタ纷换において掐蜗フレ〖ムが
 * 警ないなど·まだ评られていない眷圭は FALSE を手す. 
 * </JA>
 * <EN>
 * @brief  Compute a parameter vector from a speech window.
 *
 * This function calculates an MFCC vector from speech data windowed from
 * input speech.  The obtained MFCC vector will be stored to mfcc->tmpmfcc.
 * 
 * @param mfcc [i/o] MFCC calculation instance
 * @param window [in] speech input (windowed from input stream)
 * @param windowlen [in] length of @a window
 * 
 * @return TRUE on success (an vector obtained).  Returns FALSE if no
 * parameter vector obtained yet (due to delta delay).
 * </EN>
 *
 * @callgraph
 * @callergraph
 * 
 */
boolean
RealTimeMFCC(MFCCCalc *mfcc, SP16 *window, int windowlen)
{
  int i;
  boolean ret;
  VECT *tmpmfcc;
  Value *para;

  tmpmfcc = mfcc->tmpmfcc;
  para = mfcc->para;

  /* 不兰侨妨から base MFCC を纷换 (recog->mfccwrk を网脱) */
  /* calculate base MFCC from waveform (use recog->mfccwrk) */
  for (i=0; i < windowlen; i++) {
    mfcc->wrk->bf[i+1] = (float) window[i];
  }
  WMP_calc(mfcc->wrk, tmpmfcc, para);

  if (para->energy && para->enormal) {
    /* 滦眶エネルギ〖灌を赖惮步する */
    /* normalize log energy */
    /* リアルタイム掐蜗では券厦ごとの呵络エネルギ〖が评られないので
       木涟の券厦のパワ〖で洛脱する */
    /* Since the maximum power of the whole input utterance cannot be
       obtained at real-time input, the maximum of last input will be
       used to normalize.
    */
    tmpmfcc[para->baselen-1] = energy_max_normalize(&(mfcc->ewrk), tmpmfcc[para->baselen-1], para);
  }

  if (para->delta) {
    /* デルタを纷换する */
    /* calc delta coefficients */
    ret = WMP_deltabuf_proceed(mfcc->db, tmpmfcc);
#ifdef RDEBUG
    printf("DeltaBuf: ret=%d, status=", ret);
    for(i=0;i<mfcc->db->len;i++) {
      printf("%d", mfcc->db->is_on[i]);
    }
    printf(", nextstore=%d\n", mfcc->db->store);
#endif
    /* ret == FALSE のときはまだディレイ面なので千急借妄せず肌掐蜗へ */
    /* if ret == FALSE, there is no available frame.  So just wait for
       next input */
    if (! ret) {
      return FALSE;
    }

    /* db->vec に附哼の傅デ〖タとデルタ犯眶が掐っているので tmpmfcc にコピ〖 */
    /* now db->vec holds the current base and full delta, so copy them to tmpmfcc */
    memcpy(tmpmfcc, mfcc->db->vec, sizeof(VECT) * para->baselen * 2);
  }

  if (para->acc) {
    /* Accelerationを纷换する */
    /* calc acceleration coefficients */
    /* base+delta をそのまま掐れる */
    /* send the whole base+delta to the cycle buffer */
    ret = WMP_deltabuf_proceed(mfcc->ab, tmpmfcc);
#ifdef RDEBUG
    printf("AccelBuf: ret=%d, status=", ret);
    for(i=0;i<mfcc->ab->len;i++) {
      printf("%d", mfcc->ab->is_on[i]);
    }
    printf(", nextstore=%d\n", mfcc->ab->store);
#endif
    /* ret == FALSE のときはまだディレイ面なので千急借妄せず肌掐蜗へ */
    /* if ret == FALSE, there is no available frame.  So just wait for
       next input */
    if (! ret) {
      return FALSE;
    }
    /* ab->vec には·(base+delta) とその汗尸犯眶が掐っている. 
       [base] [delta] [delta] [acc] の界で掐っているので,
       [base] [delta] [acc] を tmpmfcc にコピ〖する. */
    /* now ab->vec holds the current (base+delta) and their delta coef. 
       it holds a vector in the order of [base] [delta] [delta] [acc], 
       so copy the [base], [delta] and [acc] to tmpmfcc.  */
    memcpy(tmpmfcc, mfcc->ab->vec, sizeof(VECT) * para->baselen * 2);
    memcpy(&(tmpmfcc[para->baselen*2]), &(mfcc->ab->vec[para->baselen*3]), sizeof(VECT) * para->baselen);
  }

#ifdef POWER_REJECT
  if (para->energy || para->c0) {
    mfcc->avg_power += tmpmfcc[para->baselen-1];
  }
#endif

  if (para->delta && (para->energy || para->c0) && para->absesup) {
    /* 冷滦猛パワ〖を近殿 */
    /* suppress absolute power */
    memmove(&(tmpmfcc[para->baselen-1]), &(tmpmfcc[para->baselen]), sizeof(VECT) * (para->vecbuflen - para->baselen));
  }

  /* この箕爬で tmpmfcc に附箕爬での呵糠の泼魔ベクトルが呈羌されている */
  /* tmpmfcc[] now holds the latest parameter vector */

  /* CMN を纷换 */
  /* perform CMN */
  if (para->cmn || para->cvn) CMN_realtime(mfcc->cmn.wrk, tmpmfcc);

  return TRUE;
}

static int
proceed_one_frame(Recog *recog)
{
  MFCCCalc *mfcc;
  RealBeam *r;
  int maxf;
  PROCESS_AM *am;
  int rewind_frame;
  boolean reprocess;
  boolean ok_p;

  r = &(recog->real);

  /* call recognition start callback */
  ok_p = FALSE;
  maxf = 0;
  for (mfcc = recog->mfcclist; mfcc; mfcc = mfcc->next) {
    if (!mfcc->valid) continue;
    if (maxf < mfcc->f) maxf = mfcc->f;
    if (mfcc->f == 0) {
      ok_p = TRUE;
    }
  }
  if (ok_p && maxf == 0) {
    /* call callback when at least one of MFCC has initial frame */
    if (recog->jconf->decodeopt.segment) {
#ifdef BACKEND_VAD
      /* not exec pass1 begin callback here */
#else
      if (!recog->process_segment) {
	callback_exec(CALLBACK_EVENT_RECOGNITION_BEGIN, recog);
      }
      callback_exec(CALLBACK_EVENT_SEGMENT_BEGIN, recog);
      callback_exec(CALLBACK_EVENT_PASS1_BEGIN, recog);
      recog->triggered = TRUE;
#endif
    } else {
      callback_exec(CALLBACK_EVENT_RECOGNITION_BEGIN, recog);
      callback_exec(CALLBACK_EVENT_PASS1_BEGIN, recog);
      recog->triggered = TRUE;
    }
  }
  /* 称インスタンスについて mfcc->f の千急借妄を1フレ〖ム渴める */
  switch (decode_proceed(recog)) {
  case -1: /* error */
    return -1;
    break;
  case 0:			/* success */
    break;
  case 1:			/* segmented */
    /* 千急借妄のセグメント妥滇で姜わったことをフラグにセット */
    /* set flag which indicates that the input has ended with segmentation request */
    r->last_is_segmented = TRUE;
    /* tell the caller to be segmented by this function */
    /* 钙び叫し傅に·ここで掐蜗を磊るよう帕える */
    return 1;
  }
#ifdef BACKEND_VAD
  /* check up trigger in case of VAD segmentation */
  if (recog->jconf->decodeopt.segment) {
    if (recog->triggered == FALSE) {
      if (spsegment_trigger_sync(recog)) {
	if (!recog->process_segment) {
	  callback_exec(CALLBACK_EVENT_RECOGNITION_BEGIN, recog);
	}
	callback_exec(CALLBACK_EVENT_SEGMENT_BEGIN, recog);
	callback_exec(CALLBACK_EVENT_PASS1_BEGIN, recog);
	recog->triggered = TRUE;
      }
    }
  }
#endif
  
  if (spsegment_need_restart(recog, &rewind_frame, &reprocess) == TRUE) {
    /* set total length to the current frame */
    for (mfcc = recog->mfcclist; mfcc; mfcc = mfcc->next) {
      if (!mfcc->valid) continue;
      mfcc->param->header.samplenum = mfcc->f + 1;
      mfcc->param->samplenum = mfcc->f + 1;
    }
    /* do rewind for all mfcc here */
    spsegment_restart_mfccs(recog, rewind_frame, reprocess);
    /* also tell adin module to rehash the concurrent audio input */
    recog->adin->rehash = TRUE;
    /* reset outprob cache for all AM */
    for(am=recog->amlist;am;am=am->next) {
      outprob_prepare(&(am->hmmwrk), am->mfcc->param->samplenum);
    }
    if (reprocess) {
      /* process the backstep MFCCs here */
      while(1) {
	ok_p = TRUE;
	for (mfcc = recog->mfcclist; mfcc; mfcc = mfcc->next) {
	  if (! mfcc->valid) continue;
	  mfcc->f++;
	  if (mfcc->f < mfcc->param->samplenum) {
	    mfcc->valid = TRUE;
	    ok_p = FALSE;
	  } else {
	    mfcc->valid = FALSE;
	  }
	}
	if (ok_p) {
	  /* すべての MFCC が姜わりに茫したのでル〖プ姜位 */
	  /* all MFCC has been processed, end of loop  */
	  for (mfcc = recog->mfcclist; mfcc; mfcc = mfcc->next) {
	    if (! mfcc->valid) continue;
	    mfcc->f--;
	  }
	  break;
	}
	/* 称インスタンスについて mfcc->f の千急借妄を1フレ〖ム渴める */
	switch (decode_proceed(recog)) {
	case -1: /* error */
	  return -1;
	  break;
	case 0:			/* success */
	  break;
	case 1:			/* segmented */
	  /* ignore segmentation while in the backstep segment */
	  break;
	}
	/* call frame-wise callback */
	callback_exec(CALLBACK_EVENT_PASS1_FRAME, recog);
      }
    }
  }
  /* call frame-wise callback if at least one of MFCC is valid at this frame */
  for (mfcc = recog->mfcclist; mfcc; mfcc = mfcc->next) {
    if (mfcc->valid) {
      callback_exec(CALLBACK_EVENT_PASS1_FRAME, recog);
      break;
    }
  }
  
  return 0;
}


/** 
 * <JA>
 * @brief  妈1パス士乖不兰千急借妄のメイン
 *
 * この簇眶柒では·敛肌弄な泼魔翁藐叫および妈1パスの千急が乖われる. 
 * 掐蜗デ〖タに滦して岭齿けˇシフトを乖いMFCC纷换を乖いながら·
 * 不兰千急を1フレ〖ムずつ事误悸乖する. 
 *
 * 千急借妄∈decode_proceed()∷において·不兰惰粗姜位が妥滇される
 * ことがある. この眷圭·踏借妄の不兰を瘦赂して妈1パスを姜位する
 * よう钙叫傅に妥滇する. 
 *
 * SPSEGMENT_NAIST あるいは GMM_VAD などのバックエンドVAD年盗箕は·デコ〖ダベ〖スの
 * VAD ∈不兰惰粗倡幌浮叫∷に燃うデコ〖ディング扩告が乖われる. 
 * トリガ涟は·千急借妄が钙ばれるが·悸狠には称簇眶柒で千急借妄は
 * 乖われていない. 倡幌を浮叫した箕·この簇眶はそこまでに评られた
 * MFCC误を办年フレ〖ム墓尸船提し·その船提し黎から奶撅の千急借妄を
 * 浩倡する. なお·剩眶借妄インスタンス粗がある眷圭·倡幌トリガは
 * どれかのインスタンスが浮叫した箕爬で链ての倡幌が票袋される. 
 * 
 * この簇眶は·不兰掐蜗ル〖チンのコ〖ルバックとして钙ばれる.
 * 不兰デ〖タの眶篱サンプル峡不ごとにこの簇眶が钙び叫される. 
 * 
 * @param Speech [in] 不兰デ〖タへのバッファへのポインタ
 * @param nowlen [in] 不兰デ〖タの墓さ
 * @param recog [i/o] engine instance
 * 
 * @return エラ〖箕に -1 を·赖撅箕に 0 を手す. また·妈1パスを
 * 姜位するよう钙叫傅に妥滇するときは 1 を手す. 
 * </JA>
 * <EN>
 * @brief  Main function of the on-the-fly 1st pass decoding
 *
 * This function performs sucessive MFCC calculation and 1st pass decoding.
 * The given input data are windowed to a certain length, then converted
 * to MFCC, and decoding for the input frame will be performed in one
 * process cycle.  The loop cycle will continue with window shift, until
 * the whole given input has been processed.
 *
 * In case of input segment request from decoding process (in
 * decode_proceed()), this function keeps the rest un-processed speech
 * to a buffer and tell the caller to stop input and end the 1st pass.
 *
 * When back-end VAD such as SPSEGMENT_NAIST or GMM_VAD is defined,  Decoder-based
 * VAD is enabled and its decoding control will be managed here.
 * In decoder-based VAD mode, the recognition will be processed but
 * no output will be done at the first un-triggering input area.
 * when speech input start is detected, this function will rewind the
 * already obtained MFCC sequence to a certain frames, and re-start
 * normal recognition at that point.  When multiple recognition process
 * instance is running, their segmentation will be synchronized.
 * 
 * This function will be called each time a new speech sample comes as
 * as callback from A/D-in routine.
 * 
 * @param Speech [in] pointer to the speech sample segments
 * @param nowlen [in] length of above
 * @param recog [i/o] engine instance
 * 
 * @return -1 on error (tell caller to terminate), 0 on success (allow caller
 * to call me for the next segment).  It returns 1 when telling the caller to
 * terminate input and go on to the next pass.
 * </EN>
 *
 * @callgraph