Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,72 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import gradio as gr
|
3 |
+
from torchaudio.sox_effects import apply_effects_file
|
4 |
+
from transformers import AutoFeatureExtractor, AutoModelForAudioXVector
|
5 |
+
|
6 |
+
device = "cuda" if toch.cuda.is_available() else "cpu"
|
7 |
+
EFFECTS = [
|
8 |
+
['remix', '-'], # pour fusionner tous les canaux
|
9 |
+
["channels", "1"], #channel-->mono
|
10 |
+
["rate", "16000"], # rééchantillonner à 16000 Hz
|
11 |
+
["gain", "-1.0"], #Atténuation -1 dB
|
12 |
+
["silence", "1", "0.1", "0.1%", "-1", "0.1", "0.1%"],
|
13 |
+
# ['pad', '0', '1.5'], # pour ajouter 1,5 seconde à la fin
|
14 |
+
['trim', '0', '10'], # obtenir les 10 premières secondes
|
15 |
+
]
|
16 |
+
|
17 |
+
model_name = "microsoft/unispeech-sat-base-plus-sv"
|
18 |
+
feature_extractor = AutoFeatureExtractor.from_pretrained(model_name)
|
19 |
+
model = AutoModelForAudioXVector.from_pretrained(model_name).to(device)
|
20 |
+
|
21 |
+
#Réglage de la valeur seuil
|
22 |
+
SEUIL = 0,85
|
23 |
+
|
24 |
+
cosine_similarity = torch.nn.CosineSimilarity(dim=-1)
|
25 |
+
|
26 |
+
def similarity_fn(path1, path2):
|
27 |
+
if not (path1 and path2):
|
28 |
+
return 'ERROR: Please record audio for *both* speakers!'
|
29 |
+
#Applying the effects to both the audio input files
|
30 |
+
wav1, _ = apply_effects_file(path1, EFFECTS)
|
31 |
+
wav2, _ = apply_effects_file(path2,EFFECTS)
|
32 |
+
#Extracting features
|
33 |
+
input1 = feature_extractor(wav1.squeeze(0), return_tensors="pt", sampling_rate=16000).input_values.to(device)
|
34 |
+
input2 = feature_extractor(wav2.squeeze(0), return_tensors="pt", sampling_rate=16000).input_values.to(device)
|
35 |
+
with torch.no_grad():
|
36 |
+
emb1 = model(input1).embeddings
|
37 |
+
emb2 = model(input2).embeddings
|
38 |
+
emb1 = torch.nn.functional.normalize(emb1, dim=-1).to(device)
|
39 |
+
emb2 = torch.nn.functional.normalize(emb2, dim=-1).to(device)
|
40 |
+
similarity = cosine_similarity(emb1, emb2).numpy()[0]
|
41 |
+
if similarity>= THRESHOLD:
|
42 |
+
return f"Similarity score is {similarity :.0%}. Audio belongs to the same person "
|
43 |
+
elif similarity< THRESHOLD:
|
44 |
+
return f"Similarity score is {similarity:.0%}. Audio doesn't belong to the same person.Authentication failed!"
|
45 |
+
|
46 |
+
inputs = [
|
47 |
+
gr.inputs.Audio(source="microphone", type="filepath", optional=True, label="Speaker #1"),
|
48 |
+
gr.inputs.Audio(source="microphone", type="filepath", optional=True, label="Speaker #2"),
|
49 |
+
]
|
50 |
+
|
51 |
+
outputs = gr.outputs.Textbox(label="Output Text")
|
52 |
+
description = (
|
53 |
+
"This app evaluates whether the given audio speech inputs belong to the same individual based on Cosine Similarity score. "
|
54 |
+
)
|
55 |
+
|
56 |
+
interface = gr.Interface(
|
57 |
+
fn=similarity_fn,
|
58 |
+
inputs=inputs,
|
59 |
+
outputs=outputs,
|
60 |
+
title="Voice Authentication with UniSpeech-SAT + X-Vectors",
|
61 |
+
description=description,
|
62 |
+
layout="horizontal",
|
63 |
+
theme="grass",
|
64 |
+
allow_flagging=False,
|
65 |
+
live=False,
|
66 |
+
examples=[
|
67 |
+
["cate_blanch.mp3", "cate_blanch_2.mp3"],
|
68 |
+
["cate_blanch.mp3", "denzel_washington.mp3"]
|
69 |
+
]
|
70 |
+
)
|
71 |
+
|
72 |
+
interface.launch(enable_queue=True)
|