Merge branch 'simplify' of...

Merge branch 'simplify' of git.zkm.de:Hertz-Lab/Research/intelligent-museum/language-identification into simplify

data/common-voice/config_cv.yaml

@@ -17,6 +17,8 @@ parallelize: True
 remove_raw: True
 
 language_table:
+  - lang: "chinese"
+    dir: "zh-CN"
   - lang: "english"
     dir: "en"
   - lang: "farsi"
@@ -25,10 +27,10 @@ language_table:
     dir: "fr"
   - lang: "german"
     dir: "de"
+  - lang: "italian"
+    dir: "it"
   - lang: "polish"
     dir: "pl"
-  - lang: "mandarin"
-    dir: "zh-CN"
   - lang: "russian"
     dir: "ru"
   - lang: "spanish"

data/common-voice/cv_to_wav.py

@@ -114,8 +114,7 @@ def traverse_csv(language, input_dir, output_dir, max_chops,
 						rand_int = np.random.randint(low=0, high=2)
 						padding_choice = ["Data", "Silence"][rand_int]
 
-
-						if vad:
+						if use_vad:
 							chips = vad.chop_from_file(wav_path_raw, padding=padding_choice)
 						else:
 							chips = chop_up_audio (wav_path_raw, padding=padding_choice,
@@ -123,8 +122,10 @@ def traverse_csv(language, input_dir, output_dir, max_chops,
 											min_length_s=min_length_s, max_silence_s=max_silence_s,
 											threshold=energy_threshold)
 
-
 						for chip_name, chip_fs, chip_data in chips:
+							if chip_data.dtype == "float32":
+								chip_data = chip_data * 32768
+								chip_data = chip_data.astype("int16")
 							wav_path = os.path.join(output_dir_wav, chip_name + ".wav")
 							wav.write(wav_path, chip_fs, chip_data)
 							produced_files += 1
@@ -155,9 +156,9 @@ if __name__ == '__main__':
 	parser = argparse.ArgumentParser()
 	parser.add_argument('--config_path', default=None,
 						help="path to the config yaml file. When given, arguments will be ignored")
-	parser.add_argument("--cv_input_dir", type=str, default=None,
+	parser.add_argument("--cv_input_dir", type=str, default="common_voice",
 						help="directory containing all languages")
-	parser.add_argument("--cv_output_dir", type=str, default="../res",
+	parser.add_argument("--cv_output_dir", type=str, default="common_voice_processed",
 						help="directory to receive converted clips of all languages")
 	# Data 
 	parser.add_argument("--max_chops", type=int, nargs=3, default=[-1, -1, -1],
@@ -189,6 +190,7 @@ if __name__ == '__main__':
 	# overwrite arguments when config is given
 	if args.config_path:
 		config = load(open(args.config_path, "rb"))
+		
 		if config is None:
 				print("Could not find config file")
 				exit(-1)
@@ -205,6 +207,7 @@ if __name__ == '__main__':
 			args.sample_width      = config["sample_width"]
 			args.parallelize       = config["parallelize"]
 			args.remove_raw        = config["remove_raw"]
+			args.use_vad           = config["use_vad"]
 			language_table         = config["language_table"]
 
 			# copy config to output dir

src/audio/chop_up.py

@@ -12,7 +12,7 @@ import os
 import time
 
 import auditok
-from .utils import pad_with_data, pad_with_noise, pad_with_silence, to_array
+from src.audio.utils import pad, to_array
 
 
 def chop_up_audio (file_name, desired_length_s = 5,
@@ -35,16 +35,8 @@ def chop_up_audio (file_name, desired_length_s = 5,
 	# extend tokens to desired length
 	audio_cuttings = []
 	for i, r in enumerate(regions):
-
 		numpy_data = to_array(r._data, 2, 1)
-
-		if padding == "Silence":
-			extended_token = pad_with_silence(numpy_data, nn_input_len)
-		elif padding == "Data":
-			extended_token = pad_with_data(numpy_data, nn_input_len)
-		else:
-			extended_token = pad_with_noise(numpy_data, nn_input_len)
-
+		extended_token = pad(numpy_data, nn_input_len, padding)
 		file_name_out = os.path.split(file_name)[-1][:-4] + "_" + str(i)
 		data_tuple = (file_name_out, sample_rate, extended_token)
 		audio_cuttings.append(data_tuple)

src/audio/silero_vad/utils_vad.py

@@ -719,7 +719,7 @@ def drop_chunks(tss: List[dict],
 
 
 import os
-from src.audio.utils import pad_with_data, pad_with_noise, pad_with_silence, to_array
+from src.audio.utils import pad
 from src.audio.utils import LogicDataSource, LogicValidater
 from auditok.core import StreamTokenizer
 
@@ -759,20 +759,11 @@ class VADTokenizer():
 		# reconstruct audio regions from index regions
 		regions = [wav[index[1]*self.vad_resolution: index[2]*self.vad_resolution] for index in indices]
 
-
 		# extend tokens to desired length
 		audio_cuttings = []
 		for i, r in enumerate(regions):
-
 			numpy_data = r.cpu().detach().numpy()
-
-			if padding == "Silence":
-				extended_token = pad_with_silence(numpy_data, self.nn_input_len)
-			elif padding == "Data":
-				extended_token = pad_with_data(numpy_data, self.nn_input_len)
-			else:
-				extended_token = pad_with_noise(numpy_data, self.nn_input_len)
-
+			extended_token = pad(numpy_data, self.nn_input_len, padding)
 			file_name_out = os.path.split(file_path)[-1][:-4] + "_" + str(i)
 			data_tuple = (file_name_out, self.sample_rate, extended_token)
 			audio_cuttings.append(data_tuple)

src/audio/utils.py

@@ -30,7 +30,7 @@ def pad_with_silence(data, max_len):
 
 def pad_with_data(data, max_len):
 	to_add = max(max_len - len(data), 0)
-	padded = np.zeros(shape=(max_len,), dtype="int16")
+	padded = np.zeros(shape=(max_len,), dtype="float32")
 	if to_add:
 		repeat = int(max_len / len(data))
 		rest = max_len % len(data)