Merge pull request #89 from valentinp72/master

Added options to bypass ffmpeg usage in some cases ; Pyro client now uses argparse

Author: DavidDoukhan

inaSpeechSegmenter/io.py

@@ -34,6 +34,26 @@ def media2sig16kmono(medianame, tmpdir=None, start_sec=None, stop_sec=None, ffmp
     Convert media to temp wav 16k mono and return signal
     """
 
+    if ffmpeg is None:
+        if start_sec is not None or stop_sec is not None:
+            raise NotImplementedError(
+                f'start_sec={start_sec} and stop_sec={stop_sec} cannot be set ' \
+                f' when running inaSpeechSegmenter without ffmpeg. Please cut '\
+                f'down your audio files beforehand or use ffmpeg.'
+            )
+        if medianame.startswith('http://') or medianame.startswith('https://'):
+            raise NotImplementedError(
+                f'Without ffmpeg you cannot process media content on http ' \
+                f'servers. You need to download your audio files beforehand ' \
+                f'or use ffmpeg. You gave medianame={medianame}.'
+            )
+
+        sig, sr = sf.read(medianame, dtype=dtype)
+        assert sr == 16_000, \
+            f'Without ffmpeg, inaSpeechSegmenter can only take files sampled ' \
+            f'at 16000 Hz. The file {medianame} is sampled at {sr} Hz.'
+        return sig
+
     base, _ = os.path.splitext(os.path.basename(medianame))
 
     with tempfile.TemporaryDirectory(dir=tmpdir) as tmpdirname:
@@ -57,4 +77,5 @@ def media2sig16kmono(medianame, tmpdir=None, start_sec=None, stop_sec=None, ffmp
         # Get Mel Power Spectrogram and Energy
         sig, sr = sf.read(tmpwav, dtype=dtype)
         assert sr == 16000
-        return sig
+        return sig
+

inaSpeechSegmenter/segmenter.py

@@ -224,9 +224,10 @@ def __init__(self, vad_engine='smn', detect_gender=True, ffmpeg='ffmpeg', batch_
                 default value (32) is slow, but works on any hardware
         """      
 
-        # test ffmpeg installation
-        if shutil.which(ffmpeg) is None:
-            raise(Exception("""ffmpeg program not found"""))
+        if ffmpeg is not None:
+            # test ffmpeg installation
+            if shutil.which(ffmpeg) is None:
+                raise(Exception("""ffmpeg program not found"""))
         self.ffmpeg = ffmpeg
 
         # set energic ratio for 1st VAD

media/musanmix.wav

@@ -79,6 +79,14 @@ def seg2str(iseg, tseg):
             self.assertEqual(curstop, nextstart,
                              '%s VS %s' % (seg2str(i, ret[i]), seg2str(i+1, ret[i+1])))
 
+        seg = Segmenter(ffmpeg=None)
+        ret = seg('./media/musanmix.wav')
+        for i in range(len(ret) -1):
+            curstop = ret[i][2]
+            nextstart = ret[i+1][1]
+            self.assertEqual(curstop, nextstart,
+                             '%s VS %s' % (seg2str(i, ret[i]), seg2str(i+1, ret[i+1])))
+
     def test_processingresult(self):
         seg = Segmenter(vad_engine='sm')
         ret = seg('./media/musanmix.mp3')
@@ -87,6 +95,14 @@ def test_processingresult(self):
         self.assertEqual([e[0] for e in ref], [e[0] for e in ret])
         np.testing.assert_almost_equal([e[1] for e in ref], [e[1] for e in ret])
         np.testing.assert_almost_equal([e[2] for e in ref], [e[2] for e in ret])
+
+        seg = Segmenter(vad_engine='sm', ffmpeg=None)
+        ret = seg('./media/musanmix.wav')
+        df = pd.read_csv('./media/musanmix-sm-gender.csv', sep='\t')
+        ref = [(l.labels, float(l.start), float(l.stop)) for _, l in df.iterrows()]
+        self.assertEqual([e[0] for e in ref], [e[0] for e in ret])
+        np.testing.assert_almost_equal([e[1] for e in ref], [e[1] for e in ret])
+        np.testing.assert_almost_equal([e[2] for e in ref], [e[2] for e in ret])
 
     def test_batch(self):
         seg = Segmenter(vad_engine='sm')
@@ -96,6 +112,12 @@ def test_batch(self):
             self.assertTrue(filecmp.cmp(lout[0], lout[1]))
             self.assertTrue(filecmp.cmp(lout[0], './media/musanmix-sm-gender.csv'))
 
+        seg = Segmenter(vad_engine='sm', ffmpeg=None)
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            lout = [os.path.join(tmpdirname, '1.1.csv'), os.path.join(tmpdirname, '2.1.csv')]
+            ret = seg.batch_process(['./media/musanmix.wav', './media/musanmix.wav'], lout)
+            self.assertTrue(filecmp.cmp(lout[0], lout[1]))
+            self.assertTrue(filecmp.cmp(lout[0], './media/musanmix-sm-gender.csv'))
 
     def test_praat_export(self):
         seg = Segmenter()

run_test.py

@@ -48,12 +48,16 @@
 parser.add_argument('-s', '--batch_size', type=int, default=32, help="(default: 32 - we recommend 1024). Size of batches to be sent to the GPU. Larger values allow faster processings, but require GPU with more memories. Default 32 size is fine even with a baseline laptop GPU.")
 parser.add_argument('-d', '--vad_engine', choices=['sm', 'smn'], default='smn', help="Voice activity detection (VAD) engine to be used (default: 'smn'). 'smn' split signal into 'speech', 'music' and 'noise' (better). 'sm' split signal into 'speech' and 'music' and do not take noise into account, which is either classified as music or speech. Results presented in ICASSP were obtained using 'sm' option")
 parser.add_argument('-g', '--detect_gender', choices = ['true', 'false'], default='True', help="(default: 'true'). If set to 'true', segments detected as speech will be splitted into 'male' and 'female' segments. If set to 'false', segments corresponding to speech will be labelled as 'speech' (faster)")
-parser.add_argument('-b', '--ffmpeg_binary', default='ffmpeg', help='Your custom binary of ffmpeg', required=False)
+parser.add_argument('-b', '--ffmpeg_binary', default='ffmpeg', help='Your custom binary of ffmpeg. Set `None` to disable ffmpeg.', required=False)
 parser.add_argument('-e', '--export_format', choices = ['csv', 'textgrid'], default='csv', help="(default: 'csv'). If set to 'csv', result will be exported in csv. If set to 'textgrid', results will be exported to praat Textgrid")
 parser.add_argument('-r', '--energy_ratio', default=0.03, type=float, help="(default: 0.03). Energetic threshold used to detect activity (percentage of mean energy of the signal)")
 
 args = parser.parse_args()
 
+if args.ffmpeg_binary.lower() == "none" or args.ffmpeg_binary == "":
+    print("Disabling ffmpeg. Make sure your audio files are already sampled at 16kHz.")
+    args.ffmpeg_binary = None
+
 # Preprocess arguments and check their consistency
 input_files = []
 for e in args.input:

scripts/ina_speech_segmenter.py

@@ -25,33 +25,50 @@
 
 
 import Pyro4
-import sys
 import os
 import socket
 
-from inaSpeechSegmenter import Segmenter
-
+import argparse
 
 if __name__ == '__main__':
-    dname = os.path.dirname(os.path.realpath(__file__))
+    parser = argparse.ArgumentParser(
+        description='Start a inaSpeechSegmenter Pyro client.'
+    )
+    parser.add_argument(
+        'uri', type=str,
+        help='URI of the Pyro server to connect and get jobs from.'
+    )
+    parser.add_argument(
+        '--batch_size', type=int, default=1024,
+        help='Batch size to use. Use lower values with small GPUs.'
+    )
+    parser.add_argument(
+        '--ffmpeg_binary', default='ffmpeg', type=str,
+        help='Your custom binary of ffmpeg. Set `None` to disable ffmpeg.'
+    )
+    args = parser.parse_args()
 
-    hostname = socket.gethostname()
+    if args.ffmpeg_binary.lower() == "none" or args.ffmpeg_binary == "":
+        print("Disabling ffmpeg. Make sure your audio files are already sampled at 16kHz.")
+        args.ffmpeg_binary = None
 
-    uri = sys.argv[1]
-    jobserver = Pyro4.Proxy(uri)
+    dname = os.path.dirname(os.path.realpath(__file__))
+    hostname = socket.gethostname()
+    jobserver = Pyro4.Proxy(args.uri)
 
     ret = -1
     outname = 'init'
-    
+
     # batch size set at 1024. Use lower values with small gpus
-    g = Segmenter(batch_size=1024)
-    
+    from inaSpeechSegmenter import Segmenter
+    g = Segmenter(batch_size=args.batch_size, ffmpeg=args.ffmpeg_binary)
+
     while True:
         lsrc, ldst = jobserver.get_njobs('%s %s' % (hostname, ret))
-            
+
         print(lsrc, ldst)
         if len(lsrc) == 0:
             print('job list finished')
             break
-        
-        ret =  g.batch_process(lsrc, ldst, skipifexist=True, nbtry=3)
+
+        ret = g.batch_process(lsrc, ldst, skipifexist=True, nbtry=3)