asahi417 commited on
Commit
e0927c5
1 Parent(s): 6ed6214
attach_speaker_embedding_s2s.py CHANGED
@@ -20,6 +20,15 @@ if se_model == "metavoice":
20
  elif se_model == "pyannote":
21
  from speaker_embedding_pyannote import PyannoteSE
22
  speaker_embedder = PyannoteSE()
 
 
 
 
 
 
 
 
 
23
  else:
24
  raise ValueError(f"unknown speaker embedding: {se_model}")
25
 
@@ -44,9 +53,14 @@ print(f"Num examples (after filtering): {len(dataset)}")
44
 
45
  def speaker_embedding(example):
46
  for side in sides:
47
- example[f"{side}.audio.speaker_embedding"] = speaker_embedder.get_speaker_embedding(
48
  example[f"{side}.audio"]["array"], example[f"{side}.audio"]["sampling_rate"]
49
  )
 
 
 
 
 
50
  return example
51
 
52
 
 
20
  elif se_model == "pyannote":
21
  from speaker_embedding_pyannote import PyannoteSE
22
  speaker_embedder = PyannoteSE()
23
+ elif se_model == "w2vbert-600m":
24
+ from speaker_embedding_hf import Wav2VecEmbedding
25
+ speaker_embedder = Wav2VecEmbedding()
26
+ elif se_model == "xlsr-2b":
27
+ from speaker_embedding_hf import XLSR2BEmbedding
28
+ speaker_embedder = XLSR2BEmbedding()
29
+ elif se_model == "hubert-xl":
30
+ from speaker_embedding_hf import HuBERTXLEmbedding
31
+ speaker_embedder = HuBERTXLEmbedding()
32
  else:
33
  raise ValueError(f"unknown speaker embedding: {se_model}")
34
 
 
53
 
54
  def speaker_embedding(example):
55
  for side in sides:
56
+ embedding = speaker_embedder.get_speaker_embedding(
57
  example[f"{side}.audio"]["array"], example[f"{side}.audio"]["sampling_rate"]
58
  )
59
+ if embedding.ndim == 1:
60
+ example[f"{side}.audio.speaker_embedding"] = embedding
61
+ else:
62
+ example[f"{side}.audio.speaker_embedding"] = embedding.mean(0)
63
+ example[f"{side}.audio.speaker_embedding.full"] = embedding
64
  return example
65
 
66
 
speaker_embedding_clap.py DELETED
@@ -1,35 +0,0 @@
1
- """CLAP embedding.
2
- - feature dimension: 512
3
- - source: https://huggingface.co/laion/larger_clap_music_and_speech
4
- """
5
- from typing import Optional
6
-
7
- import torch
8
- import librosa
9
- import numpy as np
10
- from transformers import ClapModel, ClapProcessor
11
-
12
-
13
- class ClapSE:
14
- def __init__(self, ckpt: str = "laion/larger_clap_music_and_speech"):
15
- self.model = ClapModel.from_pretrained(ckpt)
16
- self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
17
- self.model.to(self.device)
18
- self.model.eval()
19
- self.processor = ClapProcessor.from_pretrained(ckpt)
20
-
21
- def get_speaker_embedding(self, wav: np.ndarray, sampling_rate: Optional[int] = None) -> np.ndarray:
22
- if sampling_rate != self.processor.feature_extractor.sampling_rate:
23
- wav = librosa.resample(wav, orig_sr=sampling_rate, target_sr=self.processor.feature_extractor.sampling_rate)
24
- inputs = self.processor(
25
- audios=wav, sampling_rate=self.processor.feature_extractor.sampling_rate, return_tensors="pt"
26
- )
27
- with torch.no_grad():
28
- outputs = self.model.get_audio_features(**{k: v.to(self.device) for k, v in inputs.items()})
29
- return outputs.cpu().numpy()[0]
30
-
31
-
32
- class ClapGeneralSE(ClapSE):
33
-
34
- def __init__(self):
35
- super().__init__(ckpt="laion/larger_clap_general")
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
speaker_embedding_hf.py ADDED
@@ -0,0 +1,72 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Meta's w2vBERT based speaker embedding."""
2
+ from typing import Optional
3
+
4
+ import torch
5
+ import librosa
6
+ import numpy as np
7
+ from transformers import AutoModel, AutoFeatureExtractor
8
+
9
+
10
+ ############
11
+ # W2V BERT #
12
+ ############
13
+ class W2VBERTEmbedding:
14
+ def __init__(self, ckpt: str = "facebook/w2v-bert-2.0"):
15
+ self.processor = AutoFeatureExtractor.from_pretrained(ckpt)
16
+ self.model = AutoModel.from_pretrained(ckpt)
17
+ self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
18
+ self.model.to(self.device)
19
+ self.model.eval()
20
+
21
+ def get_speaker_embedding(self, wav: np.ndarray, sampling_rate: Optional[int] = None) -> np.ndarray:
22
+ # audio file is decoded on the fly
23
+ if sampling_rate != self.processor.sampling_rate:
24
+ wav = librosa.resample(wav, orig_sr=sampling_rate, target_sr=self.processor.sampling_rate)
25
+ inputs = self.processor(wav, sampling_rate=self.processor.sampling_rate, return_tensors="pt")
26
+ with torch.no_grad():
27
+ outputs = self.model(**{k: v.to(self.device) for k, v in inputs.items()})
28
+ return outputs.last_hidden_state.cpu().numpy()[0]
29
+
30
+
31
+ ##########
32
+ # HuBERT #
33
+ ##########
34
+ class HuBERTXLEmbedding(W2VBERTEmbedding):
35
+ def __init__(self):
36
+ super().__init__("facebook/hubert-xlarge-ll60k")
37
+
38
+
39
+ class HuBERTLargeEmbedding(W2VBERTEmbedding):
40
+ def __init__(self):
41
+ super().__init__("facebook/hubert-large-ll60k")
42
+
43
+
44
+ class HuBERTBaseEmbedding(W2VBERTEmbedding):
45
+ def __init__(self):
46
+ super().__init__("facebook/hubert-base-ls960")
47
+
48
+
49
+ ###########
50
+ # wav2vec #
51
+ ###########
52
+ class Wav2VecEmbedding(W2VBERTEmbedding):
53
+ def __init__(self):
54
+ super().__init__("facebook/wav2vec2-large-xlsr-53")
55
+
56
+
57
+ #########
58
+ # XLS-R #
59
+ #########
60
+ class XLSR2BEmbedding(W2VBERTEmbedding):
61
+ def __init__(self):
62
+ super().__init__("facebook/wav2vec2-xls-r-2b")
63
+
64
+
65
+ class XLSR1BEmbedding(W2VBERTEmbedding):
66
+ def __init__(self):
67
+ super().__init__("facebook/wav2vec2-xls-r-1b")
68
+
69
+
70
+ class XLSR300MEmbedding(W2VBERTEmbedding):
71
+ def __init__(self):
72
+ super().__init__("facebook/wav2vec2-xls-r-300m")