[FIX] Fix preprocessing bug for aishell3

pre.py

@@ -28,8 +28,7 @@ if __name__ == "__main__":
         "Path to the output directory that will contain the mel spectrograms, the audios and the "
         "embeds. Defaults to <datasets_root>/SV2TTS/synthesizer/")
     parser.add_argument("-n", "--n_processes", type=int, default=1, help=\
-        "Number of processes in parallel.An encoder is created for each, so you may need to lower "
+        "Number of processes in parallel.")
-        "this value on GPUs with low memory. Set it to 1 if CUDA is unhappy")
     parser.add_argument("-s", "--skip_existing", action="store_true", help=\
         "Whether to overwrite existing files with the same name. Useful if the preprocessing was "
         "interrupted. ")
@@ -40,10 +39,13 @@ if __name__ == "__main__":
     parser.add_argument("--no_alignments", action="store_true", help=\
         "Use this option when dataset does not include alignments\
         (these are used to split long audio files into sub-utterances.)")
-    parser.add_argument("--dataset", type=str, default="aidatatang_200zh", help=\
+    parser.add_argument("-d", "--dataset", type=str, default="aidatatang_200zh", help=\
         "Name of the dataset to process, allowing values: magicdata, aidatatang_200zh, aishell3.")
     parser.add_argument("-e", "--encoder_model_fpath", type=Path, default="encoder/saved_models/pretrained.pt", help=\
         "Path your trained encoder model.")
+    parser.add_argument("-ne", "--n_processes_embed", type=int, default=1, help=\
+        "Number of processes in parallel.An encoder is created for each, so you may need to lower "
+        "this value on GPUs with low memory. Set it to 1 if CUDA is unhappy")
     args = parser.parse_args()
 
     # Process the arguments
@@ -69,4 +71,4 @@ if __name__ == "__main__":
 
     preprocess_dataset(**vars(args))
     
-    create_embeddings(synthesizer_root=args.out_dir, n_processes=args.n_processes, encoder_model_fpath=encoder_model_fpath)    
+    create_embeddings(synthesizer_root=args.out_dir, n_processes=args.n_processes_embed, encoder_model_fpath=encoder_model_fpath)    

synthesizer/preprocess.py

@@ -7,6 +7,7 @@ from tqdm import tqdm
 import numpy as np
 from encoder import inference as encoder
 from synthesizer.preprocess_speaker import preprocess_speaker_general
+from synthesizer.preprocess_transcript import preprocess_transcript_aishell3
 
 data_info = {
     "aidatatang_200zh": {
@@ -22,8 +23,9 @@ data_info = {
     "aishell3":{
         "subfolders": ["train/wav"],
         "trans_filepath": "train/content.txt",
-        "speak_func": preprocess_speaker_general
+        "speak_func": preprocess_speaker_general,
-    },
+        "transcript_func": preprocess_transcript_aishell3,
+    }
 }
 
 def preprocess_dataset(datasets_root: Path, out_dir: Path, n_processes: int,
@@ -49,11 +51,15 @@ def preprocess_dataset(datasets_root: Path, out_dir: Path, n_processes: int,
     transcript_dirs = dataset_root.joinpath(dataset_info["trans_filepath"])
     assert transcript_dirs.exists(), str(transcript_dirs)+" not exist."
     with open(transcript_dirs, "r", encoding="utf-8") as dict_transcript:
-        for v in dict_transcript:
+        # process with specific function for your dataset 
-            if not v:
+        if "transcript_func" in dataset_info:
-                continue
+            dataset_info["transcript_func"](dict_info, dict_transcript)
-            v = v.strip().replace("\n","").replace("\t"," ").split(" ")
+        else:
-            dict_info[v[0]] = " ".join(v[1:])
+            for v in dict_transcript:
+                if not v:
+                    continue
+                v = v.strip().replace("\n","").replace("\t"," ").split(" ")
+                dict_info[v[0]] = " ".join(v[1:])
 
     speaker_dirs = list(chain.from_iterable(input_dir.glob("*") for input_dir in input_dirs))
     func = partial(dataset_info["speak_func"], out_dir=out_dir, skip_existing=skip_existing, 

synthesizer/preprocess_speaker.py

@@ -61,7 +61,7 @@ def _process_utterance(wav: np.ndarray, text: str, out_dir: Path, basename: str,
     return wav_fpath.name, mel_fpath.name, "embed-%s.npy" % basename, len(wav), mel_frames, text
  
 
-def _split_on_silences_aidatatang_200zh(wav_fpath, words, hparams):
+def _split_on_silences(wav_fpath, words, hparams):
     # Load the audio waveform
     wav, _ = librosa.load(wav_fpath, hparams.sample_rate)
     wav = librosa.effects.trim(wav, top_db= 40, frame_length=2048, hop_length=512)[0]
@@ -82,146 +82,18 @@ def _split_on_silences_aidatatang_200zh(wav_fpath, words, hparams):
 
 def preprocess_speaker_general(speaker_dir, out_dir: Path, skip_existing: bool, hparams, dict_info, no_alignments: bool):
     metadata = []
-    wav_fpath_list = speaker_dir.glob("*.wav")
+    extensions = ["*.wav", "*.flac", "*.mp3"]
-    # Iterate over each wav
+    for extension in extensions:
-    for wav_fpath in wav_fpath_list:
+        wav_fpath_list = speaker_dir.glob(extension)
-        words = dict_info.get(wav_fpath.name.split(".")[0])
+        # Iterate over each wav
-        words = dict_info.get(wav_fpath.name) if not words else words # try with wav 
+        for wav_fpath in wav_fpath_list:
-        if not words:
+            words = dict_info.get(wav_fpath.name.split(".")[0])
-            print("no wordS")
+            words = dict_info.get(wav_fpath.name) if not words else words # try with wav 
-            continue
+            if not words:
-        sub_basename = "%s_%02d" % (wav_fpath.name, 0)
+                print("no wordS")
-        wav, text = _split_on_silences_aidatatang_200zh(wav_fpath, words, hparams)
-        metadata.append(_process_utterance(wav, text, out_dir, sub_basename, 
-                                              skip_existing, hparams))
-    return [m for m in metadata if m is not None]
-
-def preprocess_speaker(speaker_dir, out_dir: Path, skip_existing: bool, hparams, no_alignments: bool):
-    metadata = []
-    for book_dir in speaker_dir.glob("*"):
-        if no_alignments:
-            # Gather the utterance audios and texts
-            # LibriTTS uses .wav but we will include extensions for compatibility with other datasets
-            extensions = ["*.wav", "*.flac", "*.mp3"]
-            for extension in extensions:
-                wav_fpaths = book_dir.glob(extension)
-
-                for wav_fpath in wav_fpaths:
-                    # Load the audio waveform
-                    wav, _ = librosa.load(str(wav_fpath), hparams.sample_rate)
-                    if hparams.rescale:
-                        wav = wav / np.abs(wav).max() * hparams.rescaling_max
-
-                    # Get the corresponding text
-                    # Check for .txt (for compatibility with other datasets)
-                    text_fpath = wav_fpath.with_suffix(".txt")
-                    if not text_fpath.exists():
-                        # Check for .normalized.txt (LibriTTS)
-                        text_fpath = wav_fpath.with_suffix(".normalized.txt")
-                        assert text_fpath.exists()
-                    with text_fpath.open("r") as text_file:
-                        text = "".join([line for line in text_file])
-                        text = text.replace("\"", "")
-                        text = text.strip()
-
-                    # Process the utterance
-                    metadata.append(_process_utterance(wav, text, out_dir, str(wav_fpath.with_suffix("").name),
-                                                      skip_existing, hparams))
-        else:
-            # Process alignment file (LibriSpeech support)
-            # Gather the utterance audios and texts
-            try:
-                alignments_fpath = next(book_dir.glob("*.alignment.txt"))
-                with alignments_fpath.open("r") as alignments_file:
-                    alignments = [line.rstrip().split(" ") for line in alignments_file]
-            except StopIteration:
-                # A few alignment files will be missing
                 continue
-
+            sub_basename = "%s_%02d" % (wav_fpath.name, 0)
-            # Iterate over each entry in the alignments file
+            wav, text = _split_on_silences(wav_fpath, words, hparams)
-            for wav_fname, words, end_times in alignments:
+            metadata.append(_process_utterance(wav, text, out_dir, sub_basename, 
-                wav_fpath = book_dir.joinpath(wav_fname + ".flac")
+                                                skip_existing, hparams))
-                assert wav_fpath.exists()
-                words = words.replace("\"", "").split(",")
-                end_times = list(map(float, end_times.replace("\"", "").split(",")))
-
-                # Process each sub-utterance
-                wavs, texts = _split_on_silences(wav_fpath, words, end_times, hparams)
-                for i, (wav, text) in enumerate(zip(wavs, texts)):
-                    sub_basename = "%s_%02d" % (wav_fname, i)
-                    metadata.append(_process_utterance(wav, text, out_dir, sub_basename,
-                                                      skip_existing, hparams))
-
     return [m for m in metadata if m is not None]
-
-# TODO: use original split func
-def _split_on_silences(wav_fpath, words, end_times, hparams):
-    # Load the audio waveform
-    wav, _ = librosa.load(str(wav_fpath), hparams.sample_rate)
-    if hparams.rescale:
-        wav = wav / np.abs(wav).max() * hparams.rescaling_max
-    
-    words = np.array(words)
-    start_times = np.array([0.0] + end_times[:-1])
-    end_times = np.array(end_times)
-    assert len(words) == len(end_times) == len(start_times)
-    assert words[0] == "" and words[-1] == ""
-    
-    # Find pauses that are too long
-    mask = (words == "") & (end_times - start_times >= hparams.silence_min_duration_split)
-    mask[0] = mask[-1] = True
-    breaks = np.where(mask)[0]
-
-    # Profile the noise from the silences and perform noise reduction on the waveform
-    silence_times = [[start_times[i], end_times[i]] for i in breaks]
-    silence_times = (np.array(silence_times) * hparams.sample_rate).astype(np.int)
-    noisy_wav = np.concatenate([wav[stime[0]:stime[1]] for stime in silence_times])
-    if len(noisy_wav) > hparams.sample_rate * 0.02:
-        profile = logmmse.profile_noise(noisy_wav, hparams.sample_rate)
-        wav = logmmse.denoise(wav, profile, eta=0)
-    
-    # Re-attach segments that are too short
-    segments = list(zip(breaks[:-1], breaks[1:]))
-    segment_durations = [start_times[end] - end_times[start] for start, end in segments]
-    i = 0
-    while i < len(segments) and len(segments) > 1:
-        if segment_durations[i] < hparams.utterance_min_duration:
-            # See if the segment can be re-attached with the right or the left segment
-            left_duration = float("inf") if i == 0 else segment_durations[i - 1]
-            right_duration = float("inf") if i == len(segments) - 1 else segment_durations[i + 1]
-            joined_duration = segment_durations[i] + min(left_duration, right_duration)
-
-            # Do not re-attach if it causes the joined utterance to be too long
-            if joined_duration > hparams.hop_size * hparams.max_mel_frames / hparams.sample_rate:
-                i += 1
-                continue
-
-            # Re-attach the segment with the neighbour of shortest duration
-            j = i - 1 if left_duration <= right_duration else i
-            segments[j] = (segments[j][0], segments[j + 1][1])
-            segment_durations[j] = joined_duration
-            del segments[j + 1], segment_durations[j + 1]
-        else:
-            i += 1
-    
-    # Split the utterance
-    segment_times = [[end_times[start], start_times[end]] for start, end in segments]
-    segment_times = (np.array(segment_times) * hparams.sample_rate).astype(np.int)
-    wavs = [wav[segment_time[0]:segment_time[1]] for segment_time in segment_times]
-    texts = [" ".join(words[start + 1:end]).replace("  ", " ") for start, end in segments]
-    
-    # # DEBUG: play the audio segments (run with -n=1)
-    # import sounddevice as sd
-    # if len(wavs) > 1:
-    #     print("This sentence was split in %d segments:" % len(wavs))
-    # else:
-    #     print("There are no silences long enough for this sentence to be split:")
-    # for wav, text in zip(wavs, texts):
-    #     # Pad the waveform with 1 second of silence because sounddevice tends to cut them early
-    #     # when playing them. You shouldn't need to do that in your parsers.
-    #     wav = np.concatenate((wav, [0] * 16000))
-    #     print("\t%s" % text)
-    #     sd.play(wav, 16000, blocking=True)
-    # print("")
-    
-    return wavs, texts

synthesizer/preprocess_transcript.py

@@ -0,0 +1,9 @@
+def preprocess_transcript_aishell3(dict_info, dict_transcript):
+    for v in dict_transcript:
+        if not v:
+            continue
+        v = v.strip().replace("\n","").replace("\t"," ").split(" ")
+        transList = []
+        for i in range(2, len(v), 2):
+            transList.append(v[i])
+        dict_info[v[0]] = " ".join(transList)