]> git.sesse.net Git - c64tapwav/commitdiff
Add a program to sync up two stereo channels.
authorSteinar H. Gunderson <sgunderson@bigfoot.com>
Sat, 4 May 2013 17:03:12 +0000 (19:03 +0200)
committerSteinar H. Gunderson <sgunderson@bigfoot.com>
Sat, 4 May 2013 17:03:12 +0000 (19:03 +0200)
Makefile
sync.cpp [new file with mode: 0644]

index 51c9c7c913388c5ca8f14fad73db77570e823b0d..2f1b6728df609e8472b8940abd55e0ef7d7471eb 100644 (file)
--- a/Makefile
+++ b/Makefile
@@ -1,6 +1,6 @@
 CXXFLAGS=-O2 -g -Wall
 
 CXXFLAGS=-O2 -g -Wall
 
-all: synth decode
+all: synth decode sync
 
 %.o: %.cpp
        $(CXX) -MMD -MP $(CPPFLAGS) $(CXXFLAGS) -o $@ -c $<
 
 %.o: %.cpp
        $(CXX) -MMD -MP $(CPPFLAGS) $(CXXFLAGS) -o $@ -c $<
@@ -16,5 +16,8 @@ decode: decode.o
 synth: synth.o synth_main.o
        $(CXX) -o $@ $^ $(LDFLAGS)
 
 synth: synth.o synth_main.o
        $(CXX) -o $@ $^ $(LDFLAGS)
 
+sync: sync.o
+       $(CXX) -o $@ $^ $(LDFLAGS)
+
 clean:
 clean:
-       $(RM) synth decode $(OBJS) $(DEPS)
+       $(RM) synth decode sync $(OBJS) $(DEPS)
diff --git a/sync.cpp b/sync.cpp
new file mode 100644 (file)
index 0000000..4709b84
--- /dev/null
+++ b/sync.cpp
@@ -0,0 +1,309 @@
+// A program that tries to sync up the two channels, using Viterbi decoding
+// to find the most likely misalignment as it changes throughout the file.
+
+#define NUM_THEORIES 200
+#define THEORY_FROM -10.0  /* in samples */
+#define THEORY_TO 10.0  /* in samples */
+#define SWITCH_COST 1000.0  /* pretty arbitrary */
+
+#include <stdio.h>
+#include <math.h>
+#include <vector>
+#include <algorithm>
+
+#define LANCZOS_RADIUS 30
+#define BUFSIZE 4096
+
+struct stereo_sample {
+       short left, right;
+};
+struct double_stereo_sample {
+       double left, right;
+};
+
+inline short clip(int x)
+{
+       if (x < -32768) {
+               return x;
+       } else if (x > 32767) {
+               return 32767;
+       } else {
+               return short(x);
+       }
+}
+
+double sinc(double x)
+{
+       if (fabs(x) < 1e-6) {
+               return 1.0f - fabs(x);
+       } else {
+               return sin(x) / x;
+       }
+}
+
+#if 0
+double weight(double x)
+{
+       if (fabs(x) > LANCZOS_RADIUS) {
+               return 0.0f;
+       }
+       return sinc(M_PI * x) * sinc(M_PI * x / LANCZOS_RADIUS);
+}
+
+double interpolate(const std::vector<double_stereo_sample> &pcm, double i)
+{
+       int lower = std::max<int>(ceil(i - LANCZOS_RADIUS), 0);
+       int upper = std::min<int>(floor(i + LANCZOS_RADIUS), pcm.size() - 1);
+       double sum = 0.0f;
+
+       for (int x = lower; x <= upper; ++x) {
+               sum += pcm[x].right * weight(i - x);
+       }
+       return sum;
+}
+#else
+double weight(double x)
+{
+       if (fabs(x) > 1.0f) {
+               return 0.0f;
+       }
+       return 1.0f - fabs(x);
+}
+
+inline double interpolate(const std::vector<double> &pcm, double i)
+{
+       int ii = int(i);
+       if (ii < 0 || ii >= int(pcm.size() - 1)) {
+               return 0.0;
+       }
+       double frac = i - ii;
+
+       return pcm[ii] + frac * (pcm[ii + 1] - pcm[ii]);
+}
+
+template<class T>
+inline double interpolate(const std::vector<T> &pcm, double i)
+{
+       int ii = int(i);
+       if (ii < 0 || ii >= int(pcm.size() - 1)) {
+               return 0.0;
+       }
+       double frac = i - ii;
+
+       return pcm[ii].right + frac * (pcm[ii + 1].right - pcm[ii].right);
+}
+#endif
+
+struct hypothesis {
+       int id;
+       double cost;
+       hypothesis *prev;
+};
+
+int main(int argc, char **argv)
+{
+       std::vector<stereo_sample> pcm;
+
+       while (!feof(stdin)) {
+               stereo_sample buf[BUFSIZE];
+               ssize_t ret = fread(buf, sizeof(stereo_sample), BUFSIZE, stdin);
+               if (ret >= 0) {
+                       pcm.insert(pcm.end(), buf, buf + ret);
+               }
+       }
+
+       double sum_left = 0.0, sum_right = 0.0;
+       for (unsigned i = 0; i < pcm.size(); ++i) {
+               sum_left += pcm[i].left;
+               sum_right += pcm[i].right;
+       }
+       double mean_left = sum_left / pcm.size();
+       double mean_right = sum_right / pcm.size();
+
+       //fprintf(stderr, "Mean: L=%f R=%f\n", mean_left, mean_right);
+
+       double sum2_left = 0.0, sum2_right = 0.0;
+       for (unsigned i = 0; i < pcm.size(); ++i) {
+               sum2_left += (pcm[i].left - mean_left) * (pcm[i].left - mean_left);
+               sum2_right += (pcm[i].right - mean_right) * (pcm[i].right - mean_right);
+       }
+       double var_left = sum2_left / (pcm.size() - 1);
+       double var_right = sum2_right / (pcm.size() - 1);
+
+       //fprintf(stderr, "Stddev: L=%f R=%f\n", sqrt(var_left), sqrt(var_right));
+
+       double inv_sd_left = 1.0 / sqrt(var_left);
+       double inv_sd_right = 1.0 / sqrt(var_right);
+
+       std::vector<double_stereo_sample> norm;
+       norm.resize(pcm.size());
+
+       for (unsigned i = 0; i < pcm.size(); ++i) {
+               norm[i].left = (pcm[i].left - mean_left) * inv_sd_left;
+               norm[i].right = (pcm[i].right - mean_right) * inv_sd_right;
+       }
+
+#if 0
+       double offset = 0.0;
+       double old_diff = 0.0;
+       for (unsigned i = 0; i < pcm.size(); ++i) {
+               double left = (pcm[i].left - mean_left) * inv_sd_left;
+               double right = (interpolate(pcm, i + offset) - mean_right) * inv_sd_right;
+
+               double diff = right - left;
+               old_diff = old_diff * 0.9999 + diff * 0.0001;
+               offset -= 0.1 * old_diff;
+
+               if (i % 100 == 0) {
+                       fprintf(stderr, "%7.3f: %7.3f [diff=%8.3f lagged diff=%8.3f]\n", i / 44100.0, offset, diff, old_diff);
+               }
+               printf("%f %f %f\n", i / 44100.0, left, right);
+       }
+#endif
+
+       double delays[NUM_THEORIES];
+       hypothesis *bases = new hypothesis[NUM_THEORIES];
+
+       for (int h = 0; h < NUM_THEORIES; ++h) {
+               delays[h] = THEORY_FROM + h * (THEORY_TO - THEORY_FROM) / (NUM_THEORIES - 1);
+               bases[h].id = h;
+               bases[h].cost = 0.0;
+               bases[h].prev = NULL;
+       }
+
+       hypothesis *prev_hyp = bases;
+       size_t total_end = pcm.size();
+       //size_t total_end = 4410000;
+       for (unsigned i = 0; i < total_end; i += BUFSIZE) {
+               fprintf(stderr, "%.3f\n", i / 44100.0);
+               size_t end = std::min<size_t>(i + BUFSIZE, total_end);
+       
+               hypothesis *hyp = new hypothesis[NUM_THEORIES];
+
+               // evaluate all hypotheses
+               for (int h = 0; h < NUM_THEORIES; ++h) {
+                       double sum = 0.0;
+                       double d = delays[h];
+                       for (unsigned s = i; s < end; ++s) {
+                               double left = norm[s].left;
+                               double right = interpolate(norm, s + d);
+                               double diff = (right - left) * (right - left);
+                               sum += diff;
+                       }
+
+                       double best_cost = HUGE_VAL;
+                       hypothesis *best_prev = NULL;
+                       for (int hp = 0; hp < NUM_THEORIES; ++hp) {
+                               double switch_cost = SWITCH_COST * fabs(delays[hp] - delays[h]);
+                               double cost = prev_hyp[hp].cost + sum + switch_cost;
+                               if (best_prev == NULL || cost < best_cost) {
+                                       best_cost = cost;
+                                       best_prev = &prev_hyp[hp];
+                               }
+                       }
+
+                       hyp[h].id = h;
+                       hyp[h].cost = best_cost;
+                       hyp[h].prev = best_prev;
+               }
+
+               prev_hyp = hyp;
+       }
+
+       // best winner
+       double best_cost = HUGE_VAL;
+       hypothesis *best_hyp = NULL;
+       for (int h = 0; h < NUM_THEORIES; ++h) {
+               if (best_hyp == NULL || prev_hyp[h].cost < best_cost) {
+                       best_cost = prev_hyp[h].cost;
+                       best_hyp = &prev_hyp[h];
+               }
+       }
+
+       // trace the path backwards
+       std::vector<double> best_path;
+       while (best_hyp != NULL) {
+               best_path.push_back(delays[best_hyp->id]);
+               best_hyp = best_hyp->prev;
+       }
+
+       reverse(best_path.begin(), best_path.end());
+
+       fprintf(stderr, "Writing misalignment graphs to misalignment.plot...\n");
+       FILE *fp = fopen("misalignment.plot", "w");
+       for (unsigned i = 0; i < best_path.size(); ++i) {
+               fprintf(fp, "%f %f\n", i * BUFSIZE / 44100.0, best_path[i]);
+       }
+       fclose(fp);
+
+       // readjust and write out
+       double inv_sd = sqrt(2.0) / sqrt(var_left + var_right);
+       std::vector<stereo_sample> aligned_pcm;
+       std::vector<short> mono_pcm;
+       for (unsigned i = 0; i < total_end; ++i) {
+               double d = interpolate(best_path, i / double(BUFSIZE));
+               int left = pcm[i].left;
+               int right = interpolate(pcm, i + d);
+
+               stereo_sample ss;
+               ss.left = left;
+               ss.right = clip(right);
+               aligned_pcm.push_back(ss);
+
+               mono_pcm.push_back(clip(lrintf(inv_sd * 4096.0 * (left + right))));
+       }
+
+       fprintf(stderr, "Writing realigned stereo channels to aligned.raw...\n");
+       fp = fopen("aligned.raw", "wb");
+       fwrite(aligned_pcm.data(), sizeof(stereo_sample) * aligned_pcm.size(), 1, fp);
+       fclose(fp);
+
+       fprintf(stderr, "Writing combined mono track to combined.raw...\n");
+       fp = fopen("combined.raw", "wb");
+       fwrite(mono_pcm.data(), sizeof(short) * mono_pcm.size(), 1, fp);
+       fclose(fp);
+
+       fprintf(stderr, "All done.\n");
+
+#if 0
+
+       for (int sec = 0; sec < 2400; ++sec) {
+               double from_offset = -128.0;
+               double to_offset = 128.0;
+               double best_offset = HUGE_VAL;
+               for (int iter = 0; iter < 5; ++iter) {  
+                       //printf("from %f to %f\n", from_offset, to_offset);
+                       double best_sum = HUGE_VAL;
+               
+                       for (int k = 0; k < 100; ++k) {
+                               double offset = from_offset + k * 0.01 * (to_offset - from_offset);
+                               double sum = 0.0;
+                               for (unsigned i = sec * 4410; i < sec * 4410 + 4410; ++i) {
+                               //for (unsigned i = 100000; i < 10NUM_THEORIES0; ++i) {
+                                       double left = norm[i].left;
+                                       double right = interpolate(norm, i + offset);
+                                       //double right = (norm[i + offset].right - mean_right) * inv_sd_right;
+                                       double diff = (right - left) * (right - left);
+                                       sum += diff;
+                               }
+               
+                               if (sum < best_sum) {
+                                       best_sum = sum;
+                                       best_offset = offset;
+                               }
+
+                               //printf("%f %f\n", offset, sqrt(sum / NUM_THEORIES00.0f));
+                               //fflush(stdout);
+                       }
+               
+                       double range = 0.1 * (to_offset - from_offset);
+                       from_offset = best_offset - range;
+                       to_offset = best_offset + range;
+               }
+               printf("%d %f\n", sec, best_offset);
+               fflush(stdout);
+       }
+#endif
+       
+
+}