speech-test commited on
Commit
7275eb6
β€’
1 Parent(s): 420400a
Files changed (1) hide show
  1. app.py +31 -9
app.py CHANGED
@@ -5,12 +5,25 @@ from transformers import AutoFeatureExtractor, AutoModelForAudioXVector
5
 
6
  device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
7
 
8
- OUTPUT = """
9
- <link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/css/bootstrap.min.css" integrity="sha256-YvdLHPgkqJ8DVUxjjnGVlMMJtNimJ6dYkowFFvp4kKs=" crossorigin="anonymous">
 
 
 
 
 
 
 
 
 
 
 
10
  <div class="container">
11
  <div class="row"><h1 style="text-align: center">The speakers are</h1></div>
12
- <div class="row"><h1 class="display-1" style="text-align: center">{:.1f}%</h1></div>
13
  <div class="row"><h1 style="text-align: center">similar</h1></div>
 
 
14
  </div>
15
  """
16
 
@@ -21,7 +34,9 @@ EFFECTS = [
21
  ["silence", "1", "0.1", "0.1%", "-1", "0.1", "0.1%"],
22
  ]
23
 
24
- model_name = "anton-l/unispeech-sat-base-plus-sv"
 
 
25
  feature_extractor = AutoFeatureExtractor.from_pretrained(model_name)
26
  model = AutoModelForAudioXVector.from_pretrained(model_name).to(device)
27
  cosine_sim = torch.nn.CosineSimilarity(dim=-1)
@@ -33,9 +48,10 @@ def similarity_fn(mic_path1, file_path1, mic_path2, file_path2):
33
 
34
  wav1, _ = apply_effects_file(mic_path1 if mic_path1 else file_path1, EFFECTS)
35
  wav2, _ = apply_effects_file(mic_path2 if mic_path2 else file_path2, EFFECTS)
 
36
 
37
- input1 = feature_extractor(wav1.squeeze(0), return_tensors="pt").input_values.to(device)
38
- input2 = feature_extractor(wav2.squeeze(0), return_tensors="pt").input_values.to(device)
39
 
40
  with torch.no_grad():
41
  emb1 = model(input1).embeddings
@@ -44,7 +60,12 @@ def similarity_fn(mic_path1, file_path1, mic_path2, file_path2):
44
  emb2 = torch.nn.functional.normalize(emb2, dim=-1).cpu()
45
  similarity = cosine_sim(emb1, emb2).numpy()[0]
46
 
47
- return OUTPUT.format(similarity * 100)
 
 
 
 
 
48
 
49
 
50
  inputs = [
@@ -63,7 +84,8 @@ description = (
63
  article = (
64
  "<p style='text-align: center'>"
65
  "<a href='https://huggingface.co/microsoft/unispeech-sat-large' target='_blank'>πŸŽ™οΈ Learn more about UniSpeech-SAT</a> | "
66
- "<a href='https://arxiv.org/abs/2110.05752' target='_blank'>πŸ“š Article on ArXiv</a>"
 
67
  "</p>"
68
  )
69
 
@@ -71,7 +93,7 @@ interface = gr.Interface(
71
  fn=similarity_fn,
72
  inputs=inputs,
73
  outputs=output,
74
- title="Speaker Verification with UniSpeech-SAT",
75
  description=description,
76
  article=article,
77
  layout="horizontal",
 
5
 
6
  device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
7
 
8
+ STYLE = """
9
+ <link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/css/bootstrap.min.css" integrity="sha256-YvdLHPgkqJ8DVUxjjnGVlMMJtNimJ6dYkowFFvp4kKs=" crossorigin="anonymous">
10
+ """
11
+ OUTPUT_OK = STYLE + """
12
+ <div class="container">
13
+ <div class="row"><h1 style="text-align: center">The speakers are</h1></div>
14
+ <div class="row"><h1 class="display-1 text-success" style="text-align: center">{:.1f}%</h1></div>
15
+ <div class="row"><h1 style="text-align: center">similar</h1></div>
16
+ <div class="row"><h1 class="text-success" style="text-align: center">Welcome, human!</h1></div>
17
+ <div class="row"><small style="text-align: center">(You must get 89% or more to be considered the same person)</small><div class="row">
18
+ </div>
19
+ """
20
+ OUTPUT_FAIL = STYLE + """
21
  <div class="container">
22
  <div class="row"><h1 style="text-align: center">The speakers are</h1></div>
23
+ <div class="row"><h1 class="display-1 text-danger" style="text-align: center">{:.1f}%</h1></div>
24
  <div class="row"><h1 style="text-align: center">similar</h1></div>
25
+ <div class="row"><h1 class="text-danger" style="text-align: center">You shall not pass!</h1></div>
26
+ <div class="row"><small style="text-align: center">(You must get 89% or more to be considered the same person)</small><div class="row">
27
  </div>
28
  """
29
 
 
34
  ["silence", "1", "0.1", "0.1%", "-1", "0.1", "0.1%"],
35
  ]
36
 
37
+ THRESHOLD = 0.89
38
+
39
+ model_name = "microsoft/unispeech-sat-base-plus-sv"
40
  feature_extractor = AutoFeatureExtractor.from_pretrained(model_name)
41
  model = AutoModelForAudioXVector.from_pretrained(model_name).to(device)
42
  cosine_sim = torch.nn.CosineSimilarity(dim=-1)
 
48
 
49
  wav1, _ = apply_effects_file(mic_path1 if mic_path1 else file_path1, EFFECTS)
50
  wav2, _ = apply_effects_file(mic_path2 if mic_path2 else file_path2, EFFECTS)
51
+ print(wav1.shape, wav2.shape)
52
 
53
+ input1 = feature_extractor(wav1.squeeze(0), return_tensors="pt", sampling_rate=16000).input_values.to(device)
54
+ input2 = feature_extractor(wav2.squeeze(0), return_tensors="pt", sampling_rate=16000).input_values.to(device)
55
 
56
  with torch.no_grad():
57
  emb1 = model(input1).embeddings
 
60
  emb2 = torch.nn.functional.normalize(emb2, dim=-1).cpu()
61
  similarity = cosine_sim(emb1, emb2).numpy()[0]
62
 
63
+ if similarity >= THRESHOLD:
64
+ output = OUTPUT_OK.format(similarity * 100)
65
+ else:
66
+ output = OUTPUT_FAIL.format(similarity * 100)
67
+
68
+ return output
69
 
70
 
71
  inputs = [
 
84
  article = (
85
  "<p style='text-align: center'>"
86
  "<a href='https://huggingface.co/microsoft/unispeech-sat-large' target='_blank'>πŸŽ™οΈ Learn more about UniSpeech-SAT</a> | "
87
+ "<a href='https://arxiv.org/abs/2110.05752' target='_blank'>πŸ“š UniSpeech-SAT paper</a> | "
88
+ "<a href='https://www.danielpovey.com/files/2018_icassp_xvectors.pdf' target='_blank'>πŸ“š X-Vector paper</a>"
89
  "</p>"
90
  )
91
 
 
93
  fn=similarity_fn,
94
  inputs=inputs,
95
  outputs=output,
96
+ title="Speaker Verification with UniSpeech-SAT + X-Vectors",
97
  description=description,
98
  article=article,
99
  layout="horizontal",