python/demos/demo_timestretch.py: add timescale algorithm (online version)
authorPaul Brossier <piem@piem.org>
Mon, 6 Jun 2016 13:51:16 +0000 (15:51 +0200)
committerPaul Brossier <piem@piem.org>
Mon, 6 Jun 2016 13:51:16 +0000 (15:51 +0200)
python/demos/demo_timestretch_online.py [new file with mode: 0755]

diff --git a/python/demos/demo_timestretch_online.py b/python/demos/demo_timestretch_online.py
new file mode 100755 (executable)
index 0000000..e682e8e
--- /dev/null
@@ -0,0 +1,110 @@
+#! /usr/bin/env python
+
+# Implementation of the timescale algorithm according to Dan Ellis, *A Phase
+# Vocoder in Matlab*.  http://www.ee.columbia.edu/~dpwe/resources/matlab/pvoc/
+
+# This file performs both analysis and synthesis in a single pass. See also
+# `demo_timestretch.py` for a version following the original implementation.
+
+import sys
+from aubio import source, sink, pvoc, mfcc, cvec
+from aubio import unwrap2pi, float_type
+import numpy as np
+
+win_s = 1024
+hop_s = win_s / 8 # 87.5 % overlap
+
+warmup = win_s // hop_s - 1
+
+if len(sys.argv) < 3:
+    print("Usage: %s <source_filename> <output_filename> <rate> [samplerate]".format(sys.argv[0]))
+    print("""Examples:
+    # twice faster
+    {0} track_01.mp3 track_01_faster.wav 2.0
+    # twice slower
+    {0} track_02.flac track_02_slower.wav 0.5
+    # one and a half time faster, resampling first the input to 22050
+    {0} track_02.flac track_02_slower.wav 1.5 22050""".format(sys.argv[0]))
+    sys.exit(1)
+
+source_filename = sys.argv[1]
+output_filename = sys.argv[2]
+rate = float(sys.argv[3])
+
+samplerate = 0 if len(sys.argv) < 5 else int(sys.argv[4])
+source_in = source(source_filename, samplerate, hop_s)
+samplerate = source_in.samplerate
+p = pvoc(win_s, hop_s)
+
+sink_out = sink(output_filename, samplerate)
+
+# excepted phase advance in each bin
+phi_advance = np.linspace(0, np.pi * hop_s, win_s / 2 + 1).astype (float_type)
+
+old_grain = cvec(win_s)
+new_grain = cvec(win_s)
+
+block_read = 0
+interp_read = 0
+interp_block = 0
+while True:
+
+    samples, read = source_in()
+    cur_grain = p(samples)
+
+    if block_read == 1:
+        phas_acc = old_grain.phas
+
+    #print "block_read", block_read
+    while True and (block_read > 0):
+        if interp_read >= block_read:
+            break
+        #print "`--- interp_block:", interp_block,
+        #print 'at orig_block', interp_read, '<- from', block_read - 1, block_read,
+        #print 'old_grain', old_grain, 'cur_grain', cur_grain
+        # time to compute interp grain
+        frac = 1. - np.mod(interp_read, 1.0)
+
+        # compute interpolated frame
+        new_grain.norm = frac * old_grain.norm + (1. - frac) * cur_grain.norm
+        new_grain.phas = phas_acc
+
+        # psola
+        samples = p.rdo(new_grain)
+        if interp_read > warmup: # skip the first frames to warm up phase vocoder
+            # write to sink
+            sink_out(samples, hop_s)
+
+        # calculate phase advance
+        dphas = cur_grain.phas - old_grain.phas - phi_advance
+        # unwrap angle to [-pi; pi]
+        dphas = unwrap2pi(dphas)
+        # cumulate phase, to be used for next frame
+        phas_acc += phi_advance + dphas
+
+        # prepare for next interp block
+        interp_block += 1
+        interp_read = interp_block * rate
+        if interp_read >= block_read:
+            break
+
+    # copy cur_grain to old_grain
+    old_grain.norm = np.copy(cur_grain.norm)
+    old_grain.phas = np.copy(cur_grain.phas)
+
+    block_read += 1
+    if read < hop_s: break
+
+for t in range(warmup + 2): # purge the last frames from the phase vocoder
+    new_grain.norm[:] = 0
+    new_grain.phas[:] = 0
+    samples = p.rdo(new_grain)
+    sink_out(samples, read if t == warmup + 1 else hop_s)
+
+# just to make sure
+source_in.close()
+sink_out.close()
+
+format_out = "read {:d} blocks from {:s} at {:d}Hz and rate {:f}, wrote {:d} blocks to {:s}"
+print (format_out.format(block_read, source_filename, samplerate, rate,
+    interp_block, output_filename))