# HG changeset patch # User Chris Cannam # Date 1234195532 0 # Node ID 833ca65b08203373c35a57acf4220fde11cf8dcb # Parent 09bceb0aeff641691e5f93305c94975d5fa12a9d * Update with fixes from Matthew's newer code diff -r 09bceb0aeff6 -r 833ca65b0820 dsp/tempotracking/TempoTrackV2.cpp --- a/dsp/tempotracking/TempoTrackV2.cpp Tue Jan 20 15:01:01 2009 +0000 +++ b/dsp/tempotracking/TempoTrackV2.cpp Mon Feb 09 16:05:32 2009 +0000 @@ -12,6 +12,7 @@ #include #include +#include //#define FRAMESIZE 512 @@ -25,543 +26,494 @@ void TempoTrackV2::adapt_thresh(d_vec_t &df) { + d_vec_t smoothed(df.size()); + + int p_post = 7; + int p_pre = 8; - d_vec_t smoothed(df.size()); - - int p_post = 7; - int p_pre = 8; + int t = std::min(static_cast(df.size()),p_post); // what is smaller, p_post of df size. This is to avoid accessing outside of arrays - int t = std::min(static_cast(df.size()),p_post); // what is smaller, p_post of df size. This is to avoid accessing outside of arrays + // find threshold for first 't' samples, where a full average cannot be computed yet + for (int i = 0;i <= t;i++) + { + int k = std::min((i+p_pre),static_cast(df.size())); + smoothed[i] = mean_array(df,1,k); + } + // find threshold for bulk of samples across a moving average from [i-p_pre,i+p_post] + for (uint i = t+1;i < df.size()-p_post;i++) + { + smoothed[i] = mean_array(df,i-p_pre,i+p_post); + } + // for last few samples calculate threshold, again, not enough samples to do as above + for (uint i = df.size()-p_post;i < df.size();i++) + { + int k = std::max((static_cast (i) -p_post),1); + smoothed[i] = mean_array(df,k,df.size()); + } - // find threshold for first 't' samples, where a full average cannot be computed yet - for (int i = 0;i <= t;i++) - { - int k = std::min((i+p_pre),static_cast(df.size())); - smoothed[i] = mean_array(df,1,k); - } - // find threshold for bulk of samples across a moving average from [i-p_pre,i+p_post] - for (uint i = t+1;i < df.size()-p_post;i++) - { - smoothed[i] = mean_array(df,i-p_pre,i+p_post); - } - // for last few samples calculate threshold, again, not enough samples to do as above - for (uint i = df.size()-p_post;i < df.size();i++) - { - int k = std::max((static_cast (i) -p_post),1); - smoothed[i] = mean_array(df,k,df.size()); - } - - // subtract the threshold from the detection function and check that it is not less than 0 - for (uint i = 0;i < df.size();i++) - { - df[i] -= smoothed[i]; - if (df[i] < 0) - { - df[i] = 0; - } - } + // subtract the threshold from the detection function and check that it is not less than 0 + for (uint i = 0;i < df.size();i++) + { + df[i] -= smoothed[i]; + if (df[i] < 0) + { + df[i] = 0; + } + } } double TempoTrackV2::mean_array(const d_vec_t &dfin,int start,int end) { + double sum = 0.; + + // find sum + for (int i = start;i < end;i++) + { + sum += dfin[i]; + } - double sum = 0.; - - // find sum - for (int i = start;i < end+1;i++) - { - sum += dfin[i]; - } - - return static_cast (sum / (end - start + 1) ); // average and return + return static_cast (sum / (end - start + 1) ); // average and return } void TempoTrackV2::filter_df(d_vec_t &df) { + d_vec_t a(3); + d_vec_t b(3); + d_vec_t lp_df(df.size()); + //equivalent in matlab to [b,a] = butter(2,0.4); + a[0] = 1.0000; + a[1] = -0.3695; + a[2] = 0.1958; + b[0] = 0.2066; + b[1] = 0.4131; + b[2] = 0.2066; + + double inp1 = 0.; + double inp2 = 0.; + double out1 = 0.; + double out2 = 0.; - d_vec_t a(3); - d_vec_t b(3); - d_vec_t lp_df(df.size()); - //equivalent in matlab to [b,a] = butter(2,0.4); - a[0] = 1.0000; - a[1] = -0.3695; - a[2] = 0.1958; - b[0] = 0.2066; - b[1] = 0.4131; - b[2] = 0.2066; - - double inp1 = 0.; - double inp2 = 0.; - double out1 = 0.; - double out2 = 0.; + // forwards filtering + for (uint i = 0;i < df.size();i++) + { + lp_df[i] = b[0]*df[i] + b[1]*inp1 + b[2]*inp2 - a[1]*out1 - a[2]*out2; + inp2 = inp1; + inp1 = df[i]; + out2 = out1; + out1 = lp_df[i]; + } + // copy forwards filtering to df... + // but, time-reversed, ready for backwards filtering + for (uint i = 0;i < df.size();i++) + { + df[i] = lp_df[df.size()-i-1]; + } - // forwards filtering - for (uint i = 0;i < df.size();i++) - { - lp_df[i] = b[0]*df[i] + b[1]*inp1 + b[2]*inp2 - a[1]*out1 - a[2]*out2; - inp2 = inp1; - inp1 = df[i]; - out2 = out1; - out1 = lp_df[i]; - } + for (uint i = 0;i < df.size();i++) + { + lp_df[i] = 0.; + } - - // copy forwards filtering to df... - // but, time-reversed, ready for backwards filtering - for (uint i = 0;i < df.size();i++) - { - df[i] = lp_df[df.size()-i]; - } - - for (uint i = 0;i < df.size();i++) - { - lp_df[i] = 0.; - } - - inp1 = 0.; inp2 = 0.; - out1 = 0.; out2 = 0.; + inp1 = 0.; inp2 = 0.; + out1 = 0.; out2 = 0.; // backwards filetering on time-reversed df - for (uint i = 0;i < df.size();i++) - { - lp_df[i] = b[0]*df[i] + b[1]*inp1 + b[2]*inp2 - a[1]*out1 - a[2]*out2; - inp2 = inp1; - inp1 = df[i]; - out2 = out1; - out1 = lp_df[i]; - } + for (uint i = 0;i < df.size();i++) + { + lp_df[i] = b[0]*df[i] + b[1]*inp1 + b[2]*inp2 - a[1]*out1 - a[2]*out2; + inp2 = inp1; + inp1 = df[i]; + out2 = out1; + out1 = lp_df[i]; + } // write the re-reversed (i.e. forward) version back to df - for (uint i = 0;i < df.size();i++) - { - df[i] = lp_df[df.size()-i]; - } - - + for (uint i = 0;i < df.size();i++) + { + df[i] = lp_df[df.size()-i-1]; + } } void -TempoTrackV2::calculateBeatPeriod(const d_vec_t &df, d_vec_t &beat_period) +TempoTrackV2::calculateBeatPeriod(const d_vec_t &df, d_vec_t &beat_period, + d_vec_t &tempi) { + // to follow matlab.. split into 512 sample frames with a 128 hop size + // calculate the acf, + // then the rcf.. and then stick the rcfs as columns of a matrix + // then call viterbi decoding with weight vector and transition matrix + // and get best path -// to follow matlab.. split into 512 sample frames with a 128 hop size -// calculate the acf, -// then the rcf.. and then stick the rcfs as columns of a matrix -// then call viterbi decoding with weight vector and transition matrix -// and get best path + uint wv_len = 128; + double rayparam = 43.; - uint wv_len = 128; - double rayparam = 43.; - - // make rayleigh weighting curve - d_vec_t wv(wv_len); - for (uint i=0; i (i) / pow(rayparam,2.)) * exp((-1.*pow(-static_cast (i),2.)) / (2.*pow(rayparam,2.))); - } - - - uint winlen = 512; - uint step = 128; - - d_mat_t rcfmat; - int col_counter = -1; - // main loop for beat period calculation - for (uint i=0; i<(df.size()-winlen); i+=step) - { - // get dfframe - d_vec_t dfframe(winlen); - for (uint k=0; k (i) / pow(rayparam,2.)) * exp((-1.*pow(-static_cast (i),2.)) / (2.*pow(rayparam,2.))); } - } + // beat tracking frame size (roughly 6 seconds) and hop (1.5 seconds) + uint winlen = 512; + uint step = 128; + + // matrix to store output of comb filter bank, increment column of matrix at each frame + d_mat_t rcfmat; + int col_counter = -1; + + // main loop for beat period calculation + for (uint i=0; i<(df.size()-winlen); i+=step) + { + // get dfframe + d_vec_t dfframe(winlen); + for (uint k=0; k (sum/ (dfframe.size()-lag)); + } - double sum = 0.; - double tmp = 0.; + // now apply comb filtering + int numelem = 4; + + for (uint i = 2;i < rcf.size();i++) // max beat period + { + for (int a = 1;a <= numelem;a++) // number of comb elements + { + for (int b = 1-a;b <= a-1;b++) // general state using normalisation of comb elements + { + rcf[i-1] += ( acf[(a*i+b)-1]*wv[i-1] ) / (2.*a-1.); // calculate value for comb filter row + } + } + } + + // apply adaptive threshold to rcf + adapt_thresh(rcf); + + double rcfsum =0.; + for (uint i=0; i (sum/ (dfframe.size()-lag)); - } - - -// for (uint i=0; i(i); + tmat[i][j] = exp( (-1.*pow((j-mu),2.)) / (2.*pow(sigma,2.)) ); + } + } - double sigma = 8.; - for (uint i=20;i (i); - tmat[i][j] = exp( (-1.*pow((j-mu),2.)) / (2.*pow(sigma,2.)) ); - } - } + // parameters for Viterbi decoding... this part is taken from + // Murphy's matlab - d_mat_t delta; - i_mat_t psi; - for (uint i=0;i 0 ;t--) + { + bestpath[t] = psi[t+1][bestpath[t+1]]; + } -// tmatfile.close(); + // weird but necessary hack -- couldn't get above loop to terminate at t >= 0 + bestpath[0] = psi[1][bestpath[1]]; - i_vec_t bestpath(T); - d_vec_t tmp_vec(Q); - for (uint i=0; i0 ;t--) - { - bestpath[t] = psi[t+1][bestpath[t+1]]; - } - // very weird hack! - bestpath[0] = psi[1][bestpath[1]]; + //fill in the last values... + for (uint i=lastind; i (beat_period[i]); - txwt[j] = exp( -0.5*pow(tightness * log((round(2*mu)-j)/mu),2)); - - scorecands[j] = txwt[j] * cumscore[i+prange_min+j]; + localscore[i] = df[i]; + backlink[i] = -1; } - double vv = get_max_val(scorecands); - int xx = get_max_ind(scorecands); + double tightness = 4.; + double alpha = 0.9; - cumscore[i] = alpha*vv + (1.-alpha)*localscore[i]; + // main loop + for (uint i=0; i (beat_period[i]); + txwt[j] = exp( -0.5*pow(tightness * log((round(2*mu)-j)/mu),2)); + // IF IN THE ALLOWED RANGE, THEN LOOK AT CUMSCORE[I+PRANGE_MIN+J + // ELSE LEAVE AT DEFAULT VALUE FROM INITIALISATION: D_VEC_T SCORECANDS (TXWT.SIZE()); - d_vec_t tmp_vec; - for (uint i=cumscore.size() - beat_period[beat_period.size()-1] ; i= 0) + { + scorecands[j] = txwt[j] * cumscore[cscore_ind]; + } + } - int startpoint = get_max_ind(tmp_vec) + cumscore.size() - beat_period[beat_period.size()-1] ; + // find max value and index of maximum value + double vv = get_max_val(scorecands); + int xx = get_max_ind(scorecands); - i_vec_t ibeats; - ibeats.push_back(startpoint); - while (backlink[ibeats.back()] > 3*beat_period[0]) - { - ibeats.push_back(backlink[ibeats.back()]); - } + cumscore[i] = alpha*vv + (1.-alpha)*localscore[i]; + backlink[i] = i+prange_min+xx; + } + + // STARTING POINT, I.E. LAST BEAT.. PICK A STRONG POINT IN cumscore VECTOR + d_vec_t tmp_vec; + for (uint i=cumscore.size() - beat_period[beat_period.size()-1] ; i 0) + { + ibeats.push_back(backlink[ibeats.back()]); + } - - for (uint i=0; i(ibeats[i]) ); - - // cout << ibeats[i] << " " << beats[i] <(ibeats[ibeats.size()-i-1]) ); + } } diff -r 09bceb0aeff6 -r 833ca65b0820 dsp/tempotracking/TempoTrackV2.h --- a/dsp/tempotracking/TempoTrackV2.h Tue Jan 20 15:01:01 2009 +0000 +++ b/dsp/tempotracking/TempoTrackV2.h Mon Feb 09 16:05:32 2009 +0000 @@ -23,7 +23,8 @@ ~TempoTrackV2(); void calculateBeatPeriod(const vector &df, - vector &beatPeriod); + vector &beatPeriod, + vector &tempi); void calculateBeats(const vector &df, const vector &beatPeriod, @@ -39,7 +40,8 @@ double mean_array(const d_vec_t &dfin, int start, int end); void filter_df(d_vec_t &df); void get_rcf(const d_vec_t &dfframe, const d_vec_t &wv, d_vec_t &rcf); - void viterbi_decode(const d_mat_t &rcfmat, const d_vec_t &wv, d_vec_t &bp); + void viterbi_decode(const d_mat_t &rcfmat, const d_vec_t &wv, + d_vec_t &bp, d_vec_t &tempi); double get_max_val(const d_vec_t &df); int get_max_ind(const d_vec_t &df); void normalise_vec(d_vec_t &df);