New bandwidth detection code

Can now work up to full-band and uses lsb_depth to fix the noise issue.

4 files changed, 76 insertions, 33 deletions

diff --git a/celt/celt.h b/celt/celt.h
index a8e8c7fd..614698f0 100644
--- a/celt/celt.h
+++ b/celt/celt.h
@@ -57,6 +57,8 @@ typedef struct {
    opus_val16 noisiness;
    opus_val16 activity;
    opus_val16 music_prob;
+   int        bandwidth;
+   int        opus_bandwidth;
 }AnalysisInfo;
 
 #define __celt_check_mode_ptr_ptr(ptr) ((ptr) + ((ptr) - (const CELTMode**)(ptr)))
diff --git a/src/analysis.c b/src/analysis.c
index b3be892a..2f963510 100644
--- a/src/analysis.c
+++ b/src/analysis.c
@@ -100,6 +100,10 @@ static const int tbands[NB_TBANDS+1] = {
        2,  4,  6,  8, 10, 12, 14, 16, 20, 24, 28, 32, 40, 48, 56, 68, 80, 96, 120
 };
 
+static const int extra_bands[NB_TOT_BANDS+1] = {
+      0, 2,  4,  6,  8, 10, 12, 14, 16, 20, 24, 28, 32, 40, 48, 56, 68, 80, 96, 120, 160, 200
+};
+
 /*static const float tweight[NB_TBANDS+1] = {
       .3, .4, .5, .6, .7, .8, .9, 1., 1., 1., 1., 1., 1., 1., .8, .7, .6, .5
 };*/
@@ -135,7 +139,7 @@ static inline float fast_atan2f(float y, float x) {
    }
 }
 
-void tonality_analysis(TonalityAnalysisState *tonal, AnalysisInfo *info, CELTEncoder *celt_enc, const opus_val16 *x, int C)
+void tonality_analysis(TonalityAnalysisState *tonal, AnalysisInfo *info, CELTEncoder *celt_enc, const opus_val16 *x, int C, int lsb_depth)
 {
     int i, b;
     const CELTMode *mode;
@@ -153,7 +157,7 @@ void tonality_analysis(TonalityAnalysisState *tonal, AnalysisInfo *info, CELTEnc
     float features[100];
     float frame_tonality;
     float max_frame_tonality;
-    float tw_sum=0;
+    /*float tw_sum=0;*/
     float frame_noisiness;
     const float pi4 = M_PI*M_PI*M_PI*M_PI;
     float slope=0;
@@ -164,7 +168,8 @@ void tonality_analysis(TonalityAnalysisState *tonal, AnalysisInfo *info, CELTEnc
     float frame_loudness;
     float bandwidth_mask;
     int bandwidth=0;
-    float bandE[NB_TBANDS];
+    float maxE = 0;
+    float noise_floor;
     celt_encoder_ctl(celt_enc, CELT_GET_MODE(&mode));
 
     tonal->last_transition++;
@@ -236,7 +241,7 @@ void tonality_analysis(TonalityAnalysisState *tonal, AnalysisInfo *info, CELTEnc
 
     frame_tonality = 0;
     max_frame_tonality = 0;
-    tw_sum = 0;
+    /*tw_sum = 0;*/
     info->activity = 0;
     frame_noisiness = 0;
     frame_stationarity = 0;
@@ -264,19 +269,10 @@ void tonality_analysis(TonalityAnalysisState *tonal, AnalysisInfo *info, CELTEnc
           tE += binE*tonality[i];
           nE += binE*2*(.5-noisiness[i]);
        }
-       bandE[b] = E;
        tonal->E[tonal->E_count][b] = E;
        frame_noisiness += nE/(1e-15+E);
 
        frame_loudness += sqrt(E+1e-10);
-       /* Add a reasonable noise floor */
-       tonal->meanE[b] = (1-alphaE2)*tonal->meanE[b] + alphaE2*E;
-       tonal->meanRE[b] = (1-alphaE2)*tonal->meanRE[b] + alphaE2*sqrt(E);
-       /* 13 dB slope for spreading function */
-       bandwidth_mask = MAX32(.05*bandwidth_mask, E);
-       /* Checks if band looks like stationary noise or if it's below a (trivial) masking curve */
-       if (tonal->meanRE[b]*tonal->meanRE[b] < tonal->meanE[b]*.95 && E>.1*bandwidth_mask)
-          bandwidth = b;
        logE[b] = log(E+1e-10);
        tonal->lowE[b] = MIN32(logE[b], tonal->lowE[b]+.01);
        tonal->highE[b] = MAX32(logE[b], tonal->highE[b]-.1);
@@ -317,6 +313,41 @@ void tonality_analysis(TonalityAnalysisState *tonal, AnalysisInfo *info, CELTEnc
        tonal->prev_band_tonality[b] = band_tonality[b];
     }
 
+    bandwidth_mask = 0;
+    bandwidth = 0;
+    for (b=0;b<NB_TOT_BANDS;b++)
+       maxE = MAX32(maxE, tonal->meanE[b]);
+    noise_floor = 3.2e-7f/(1<<(2*IMAX(0,lsb_depth-8)));
+    for (b=0;b<NB_TOT_BANDS;b++)
+    {
+       float E=0;
+       int band_start, band_end;
+       /* Keep a margin of 300 Hz for aliasing */
+       band_start = extra_bands[b]+3;
+       band_end = extra_bands[b+1]+3;
+       for (i=band_start;i<band_end;i++)
+       {
+          float binE = out[i].r*out[i].r + out[N-i].r*out[N-i].r
+                     + out[i].i*out[i].i + out[N-i].i*out[N-i].i;
+          E += binE;
+       }
+       E /= (band_end-band_start);
+       maxE = MAX32(maxE, E);
+       if (tonal->count>2)
+       {
+          tonal->meanE[b] = (1-alphaE2)*tonal->meanE[b] + alphaE2*E;
+       } else {
+          tonal->meanE[b] = E;
+       }
+       E = MAX32(E, tonal->meanE[b]);
+       /* 13 dB slope for spreading function */
+       bandwidth_mask = MAX32(.05*bandwidth_mask, E);
+       /* Checks if band looks like stationary noise or if it's below a (trivial) masking curve */
+       if (E>.1*bandwidth_mask && E*1e10f > maxE && E > noise_floor)
+          bandwidth = b;
+    }
+    if (tonal->count<=2)
+       bandwidth = 20;
     frame_loudness = 20*log10(frame_loudness);
     tonal->Etracker = MAX32(tonal->Etracker-.03, frame_loudness);
     tonal->lowECount *= (1-alphaE);
@@ -417,21 +448,20 @@ void tonality_analysis(TonalityAnalysisState *tonal, AnalysisInfo *info, CELTEnc
        printf("%f ", features[i]);
     printf("\n");*/
 
-    /* FIXME: Can't detect SWB for now because the last band ends at 12 kHz */
-    if (bandwidth == NB_TBANDS-1 || tonal->count<100)
-    {
+    if (bandwidth<=12 || (bandwidth==13 && tonal->opus_bandwidth == OPUS_BANDWIDTH_NARROWBAND))
+       tonal->opus_bandwidth = OPUS_BANDWIDTH_NARROWBAND;
+    else if (bandwidth<=14 || (bandwidth==15 && tonal->opus_bandwidth == OPUS_BANDWIDTH_MEDIUMBAND))
+       tonal->opus_bandwidth = OPUS_BANDWIDTH_MEDIUMBAND;
+    else if (bandwidth<=16 || (bandwidth==17 && tonal->opus_bandwidth == OPUS_BANDWIDTH_WIDEBAND))
+       tonal->opus_bandwidth = OPUS_BANDWIDTH_WIDEBAND;
+    else if (bandwidth<=18)
+       tonal->opus_bandwidth = OPUS_BANDWIDTH_SUPERWIDEBAND;
+    else
        tonal->opus_bandwidth = OPUS_BANDWIDTH_FULLBAND;
-    } else {
-       int close_enough = 0;
-       if (bandE[bandwidth-1] < 3000*bandE[NB_TBANDS-1] && bandwidth < NB_TBANDS-1)
-          close_enough=1;
-       if (bandwidth<=11 || (bandwidth==12 && close_enough))
-          tonal->opus_bandwidth = OPUS_BANDWIDTH_NARROWBAND;
-       else if (bandwidth<=13)
-          tonal->opus_bandwidth = OPUS_BANDWIDTH_MEDIUMBAND;
-       else if (bandwidth<=15 || (bandwidth==16 && close_enough))
-          tonal->opus_bandwidth = OPUS_BANDWIDTH_WIDEBAND;
-    }
+
+    info->bandwidth = bandwidth;
+    info->opus_bandwidth = tonal->opus_bandwidth;
+    /*printf("%d %d\n", info->bandwidth, info->opus_bandwidth);*/
     info->noisiness = frame_noisiness;
     info->valid = 1;
 }
diff --git a/src/analysis.h b/src/analysis.h
index 09d1036a..bf8ad40a 100644
--- a/src/analysis.h
+++ b/src/analysis.h
@@ -30,7 +30,7 @@
 
 #define NB_FRAMES 8
 #define NB_TBANDS 18
-
+#define NB_TOT_BANDS 21
 
 typedef struct {
    float angle[240];
@@ -40,7 +40,7 @@ typedef struct {
    float prev_tonality;
    float E[NB_FRAMES][NB_TBANDS];
    float lowE[NB_TBANDS], highE[NB_TBANDS];
-   float meanE[NB_TBANDS], meanRE[NB_TBANDS];
+   float meanE[NB_TOT_BANDS];
    float mem[32];
    float cmean[8];
    float std[9];
@@ -55,6 +55,6 @@ typedef struct {
 } TonalityAnalysisState;
 
 void tonality_analysis(TonalityAnalysisState *tonal, AnalysisInfo *info,
-     CELTEncoder *celt_enc, const opus_val16 *x, int C);
+     CELTEncoder *celt_enc, const opus_val16 *x, int C, int lsb_depth);
 
 #endif
diff --git a/src/opus_encoder.c b/src/opus_encoder.c
index 6c491b59..5199d664 100644
--- a/src/opus_encoder.c
+++ b/src/opus_encoder.c
@@ -69,6 +69,7 @@ struct OpusEncoder {
     int          vbr_constraint;
     opus_int32   bitrate_bps;
     opus_int32   user_bitrate_bps;
+    int          lsb_depth;
     int          encoder_buffer;
 
 #define OPUS_ENCODER_RESET_START stream_channels
@@ -210,6 +211,7 @@ int opus_encoder_init(OpusEncoder* st, opus_int32 Fs, int channels, int applicat
     st->user_forced_mode = OPUS_AUTO;
     st->voice_ratio = -1;
     st->encoder_buffer = st->Fs/100;
+    st->lsb_depth = 24;
 
     /* Delay compensation of 4 ms (2.5 ms for SILK's extra look-ahead 
        + 1.5 ms for SILK resamplers and stereo prediction) */
@@ -859,6 +861,13 @@ opus_int32 opus_encode_float(OpusEncoder *st, const opus_val16 *pcm, int frame_s
         st->bandwidth = OPUS_BANDWIDTH_MEDIUMBAND;
     if (st->Fs <= 8000 && st->bandwidth > OPUS_BANDWIDTH_NARROWBAND)
         st->bandwidth = OPUS_BANDWIDTH_NARROWBAND;
+#ifndef FIXED_POINT
+    if (analysis_info.valid)
+    {
+       st->bandwidth = IMIN(st->bandwidth, analysis_info.opus_bandwidth);
+    }
+#endif
+    celt_encoder_ctl(celt_enc, OPUS_SET_LSB_DEPTH(st->lsb_depth));
 
     /* If max_data_bytes represents less than 8 kb/s, switch to CELT-only mode */
     if (max_data_bytes < (frame_rate > 50 ? 12000 : 8000)*frame_size / (st->Fs * 8))
@@ -976,7 +985,7 @@ opus_int32 opus_encode_float(OpusEncoder *st, const opus_val16 *pcm, int frame_s
        int nb_analysis_frames;
        nb_analysis_frames = frame_size/(st->Fs/100);
        for (i=0;i<nb_analysis_frames;i++)
-          tonality_analysis(&st->analysis, &analysis_info, celt_enc, pcm_buf+i*(st->Fs/100)*st->channels, st->channels);
+          tonality_analysis(&st->analysis, &analysis_info, celt_enc, pcm_buf+i*(st->Fs/100)*st->channels, st->channels, st->lsb_depth);
        if (st->signal_type == OPUS_AUTO)
           st->voice_ratio = (int)floor(.5+100*(1-analysis_info.music_prob));
     } else {
@@ -1700,13 +1709,15 @@ int opus_encoder_ctl(OpusEncoder *st, int request, ...)
         case OPUS_SET_LSB_DEPTH_REQUEST:
         {
             opus_int32 value = va_arg(ap, opus_int32);
-            ret = celt_encoder_ctl(celt_enc, OPUS_SET_LSB_DEPTH(value));
+            if (value<8 || value>24)
+               goto bad_arg;
+            st->lsb_depth=value;
         }
         break;
         case OPUS_GET_LSB_DEPTH_REQUEST:
         {
             opus_int32 *value = va_arg(ap, opus_int32*);
-            celt_encoder_ctl(celt_enc, OPUS_GET_LSB_DEPTH(value));
+            *value = st->lsb_depth;
         }
         break;
         case OPUS_RESET_STATE: