asahi417 commited on
Commit
4578bc8
·
1 Parent(s): 2ace478
attach_speaker_embedding_s2s.py ADDED
@@ -0,0 +1,156 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import subprocess
2
+ from os.path import join as p_join
3
+ from typing import Optional
4
+
5
+ import librosa
6
+ from librosa import feature
7
+ import numpy as np
8
+ from torch import nn
9
+
10
+
11
+ import os
12
+ from os.path import expanduser
13
+
14
+ import shutil
15
+ import torch
16
+ from soundfile import LibsndfileError
17
+ from datasets import load_dataset, DatasetDict, Audio
18
+
19
+ direction = os.getenv("DIRECTION", "enA-jaA")
20
+ sides = set(direction.split("-"))
21
+ dataset_id = os.getenv("DATASET_ID", 0)
22
+ num_proc = int(os.getenv("NUM_PROC", 1))
23
+ hf_org = os.getenv("HF_ORG", "asahi417")
24
+ hf_dataset = f"seamless-align-{direction}"
25
+ dataset = load_dataset(f"{hf_org}/{hf_dataset}", f"subset_{dataset_id}", split="train")
26
+ audio_loader = Audio()
27
+
28
+ checkpoint_url = "https://huggingface.co/datasets/asahi417/experiment-speaker-embedding/resolve/main/meta_voice_speaker_encoder.pt"
29
+ model_weight = p_join(os.path.expanduser('~'), ".cache", "experiment_speaker_embedding", "meta_voice_speaker_encoder.pt")
30
+
31
+
32
+ def wget(url: str, output_file: Optional[str] = None):
33
+ os.makedirs(os.path.dirname(output_file), exist_ok=True)
34
+ subprocess.run(["wget", url, "-O", output_file])
35
+ if not os.path.exists(output_file):
36
+ raise ValueError(f"failed to download {url}")
37
+
38
+
39
+ class MetaVoiceSE(nn.Module):
40
+
41
+ mel_window_length = 25
42
+ mel_window_step = 10
43
+ mel_n_channels = 40
44
+ sampling_rate = 16000
45
+ partials_n_frames = 160
46
+ model_hidden_size = 256
47
+ model_embedding_size = 256
48
+ model_num_layers = 3
49
+
50
+ def __init__(self):
51
+ super().__init__()
52
+ if not os.path.exists(model_weight):
53
+ wget(checkpoint_url, model_weight)
54
+ # Define the network
55
+ self.lstm = nn.LSTM(self.mel_n_channels, self.model_hidden_size, self.model_num_layers, batch_first=True)
56
+ self.linear = nn.Linear(self.model_hidden_size, self.model_embedding_size)
57
+ self.relu = nn.ReLU()
58
+ # Load weight
59
+ self.load_state_dict(torch.load(model_weight, map_location="cpu")["model_state"], strict=False)
60
+ # Get the target device
61
+ self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
62
+ self.to(self.device)
63
+ self.eval()
64
+
65
+ def compute_partial_slices(self, n_samples: int, rate, min_coverage):
66
+ # Compute how many frames separate two partial utterances
67
+ samples_per_frame = int((self.sampling_rate * self.mel_window_step / 1000))
68
+ n_frames = int(np.ceil((n_samples + 1) / samples_per_frame))
69
+ frame_step = int(np.round((self.sampling_rate / rate) / samples_per_frame))
70
+ # Compute the slices
71
+ wav_slices, mel_slices = [], []
72
+ steps = max(1, n_frames - self.partials_n_frames + frame_step + 1)
73
+ for i in range(0, steps, frame_step):
74
+ mel_range = np.array([i, i + self.partials_n_frames])
75
+ wav_range = mel_range * samples_per_frame
76
+ mel_slices.append(slice(*mel_range))
77
+ wav_slices.append(slice(*wav_range))
78
+ # Evaluate whether extra padding is warranted or not
79
+ last_wav_range = wav_slices[-1]
80
+ coverage = (n_samples - last_wav_range.start) / (last_wav_range.stop - last_wav_range.start)
81
+ if coverage < min_coverage and len(mel_slices) > 1:
82
+ return wav_slices[:-1], mel_slices[:-1]
83
+ return wav_slices, mel_slices
84
+
85
+ def get_speaker_embedding(self,
86
+ wav: np.ndarray,
87
+ sampling_rate: Optional[int] = None,
88
+ rate: float = 1.3,
89
+ min_coverage: float = 0.75) -> np.ndarray:
90
+ if sampling_rate != self.sampling_rate:
91
+ wav = librosa.resample(wav, orig_sr=sampling_rate, target_sr=self.sampling_rate)
92
+ wav, _ = librosa.effects.trim(wav, top_db=20)
93
+ wav_slices, mel_slices = self.compute_partial_slices(len(wav), rate, min_coverage)
94
+ max_wave_length = wav_slices[-1].stop
95
+ if max_wave_length >= len(wav):
96
+ wav = np.pad(wav, (0, max_wave_length - len(wav)), "constant")
97
+ # Wav -> Mel spectrogram
98
+ frames = feature.melspectrogram(
99
+ y=wav,
100
+ sr=self.sampling_rate,
101
+ n_fft=int(self.sampling_rate * self.mel_window_length / 1000),
102
+ hop_length=int(self.sampling_rate * self.mel_window_step / 1000),
103
+ n_mels=self.mel_n_channels,
104
+ )
105
+ mel = frames.astype(np.float32).T
106
+ mel = np.array([mel[s] for s in mel_slices])
107
+ # inference
108
+ with torch.no_grad():
109
+ mel = torch.from_numpy(mel).to(self.device)
110
+ _, (hidden, _) = self.lstm(mel)
111
+ embeds_raw = self.relu(self.linear(hidden[-1]))
112
+ partial_embeds = embeds_raw / torch.norm(embeds_raw, dim=1, keepdim=True)
113
+ partial_embeds = partial_embeds.cpu().numpy()
114
+ raw_embed = np.mean(partial_embeds, axis=0)
115
+ return raw_embed / np.linalg.norm(raw_embed, 2)
116
+
117
+
118
+ speaker_embedder = MetaVoiceSE()
119
+
120
+
121
+ def error_file(example):
122
+ for side in sides:
123
+ try:
124
+ audio_loader.decode_example(example[f"{side}.audio"])
125
+ except LibsndfileError:
126
+ return False
127
+ return True
128
+
129
+
130
+ print(f"Num examples: {len(dataset)}")
131
+ for s in sides:
132
+ dataset = dataset.cast_column(f"{s}.audio", Audio(decode=False))
133
+ dataset = dataset.filter(error_file, num_proc=num_proc, desc="drop broken audio")
134
+ for s in sides:
135
+ dataset = dataset.cast_column(f"{s}.audio", Audio())
136
+ print(f"Num examples (after filtering): {len(dataset)}")
137
+
138
+
139
+ def speaker_embedding(example):
140
+ for side in sides:
141
+ example[f"{side}.audio.speaker_embedding"] = speaker_embedder.get_speaker_embedding(
142
+ example[f"{side}.audio"]["array"], example[f"{side}.audio"]["sampling_rate"]
143
+ )
144
+ return example
145
+
146
+
147
+ dataset = dataset.map(
148
+ function=speaker_embedding,
149
+ remove_columns=[f"{s}.audio" for s in sides] + [f"{s}.url" for s in sides] + [f"{s}.duration_start" for s in sides] + [f"{s}.duration_end" for s in sides],
150
+ num_proc=num_proc,
151
+ desc="attach speaker embedding dataset"
152
+ )
153
+ DatasetDict({"train": dataset}).push_to_hub(f"{hf_org}/{hf_dataset}.speaker-embedding.metavoice", config_name=f"subset_{dataset_id}")
154
+ cache_dir = f"{expanduser('~')}/.cache/huggingface/datasets/{hf_org}___{hf_dataset}/subset_{dataset_id}"
155
+ if os.path.exists(cache_dir):
156
+ shutil.rmtree(cache_dir)
main_s2s.sh CHANGED
@@ -64,7 +64,7 @@ export LINE_NO_START=0
64
  export LINE_NO_END=10
65
  python fetch_dataset_s2s.py
66
  # main
67
- for i in $(seq 121 149);
68
  do
69
  export N_POOL=15
70
  export DATASET_ID=${i}
@@ -85,7 +85,7 @@ do
85
  python fetch_dataset_s2s.py
86
  done
87
  # tokenize
88
- for i in $(seq 50 120);
89
  do
90
  export DATASET_ID=${i}
91
  export DIRECTION="enA-viA"
@@ -201,3 +201,24 @@ export DIRECTION="deA-enA"
201
  export LINE_NO_START=0
202
  export LINE_NO_END=10
203
  python fetch_dataset_s2s.py
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
64
  export LINE_NO_END=10
65
  python fetch_dataset_s2s.py
66
  # main
67
+ for i in $(seq 1 149);
68
  do
69
  export N_POOL=15
70
  export DATASET_ID=${i}
 
85
  python fetch_dataset_s2s.py
86
  done
87
  # tokenize
88
+ for i in $(seq 120 140);
89
  do
90
  export DATASET_ID=${i}
91
  export DIRECTION="enA-viA"
 
201
  export LINE_NO_START=0
202
  export LINE_NO_END=10
203
  python fetch_dataset_s2s.py
204
+ # main
205
+ for i in $(seq 1 200);
206
+ do
207
+ export N_POOL=15
208
+ export DATASET_ID=${i}
209
+ export DIRECTION="deA-enA"
210
+ export LINE_NO_START=$(((DATASET_ID-1) * 2500))
211
+ export LINE_NO_END=$((DATASET_ID * 2500))
212
+ echo ${LINE_NO_START}
213
+ python fetch_dataset_s2s.py
214
+ done
215
+ for i in $(seq 201 394);
216
+ do
217
+ export N_POOL=15
218
+ export DATASET_ID=${i}
219
+ export DIRECTION="deA-enA"
220
+ export LINE_NO_START=$(((DATASET_ID-1) * 2500))
221
+ export LINE_NO_END=$((DATASET_ID * 2500))
222
+ echo ${LINE_NO_START}
223
+ python fetch_dataset_s2s.py
224
+ done
tokenize_dataset_s2s.py CHANGED
@@ -11,7 +11,6 @@ from encodec_audio_tokenizer import EncodecTokenizer
11
  direction = os.getenv("DIRECTION", "enA-jaA")
12
  sides = set(direction.split("-"))
13
  dataset_id = os.getenv("DATASET_ID", 0)
14
- batch_size = int(os.getenv("BATCH_SIZE", 64))
15
  num_proc = int(os.getenv("NUM_PROC", 1))
16
  hf_org = os.getenv("HF_ORG", "asahi417")
17
  hf_dataset = f"seamless-align-{direction}"
 
11
  direction = os.getenv("DIRECTION", "enA-jaA")
12
  sides = set(direction.split("-"))
13
  dataset_id = os.getenv("DATASET_ID", 0)
 
14
  num_proc = int(os.getenv("NUM_PROC", 1))
15
  hf_org = os.getenv("HF_ORG", "asahi417")
16
  hf_dataset = f"seamless-align-{direction}"