aliasgerovs's picture
Update app.py
a16ea6c verified
raw
history blame
6.18 kB
import gradio as gr
import torch
import torchaudio
from speechbrain.pretrained import SpeakerRecognition
import torch.nn as nn
from transformers import AutoModel
import os
from huggingface_hub import hf_hub_download
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
speaker_model = SpeakerRecognition.from_hparams(
source="speechbrain/spkrec-ecapa-voxceleb",
savedir="tmp",
run_opts={"device": device}
)
class PretrainedTransformerClassifier(nn.Module):
def __init__(self, num_classes=3):
super().__init__()
self.transformer = AutoModel.from_pretrained('distilbert/distilroberta-base')
for param in self.transformer.parameters():
param.requires_grad = False
for param in self.transformer.encoder.layer[-2:].parameters():
param.requires_grad = True
self.embed_projection = nn.Sequential(
nn.Linear(1, 768),
nn.LayerNorm(768),
nn.Dropout(0.1)
)
self.classifier = nn.Sequential(
nn.Linear(768, 256),
nn.ReLU(),
nn.Dropout(0.1),
nn.Linear(256, num_classes)
)
def forward(self, x):
x = self.embed_projection(x.unsqueeze(-1))
if len(x.shape) == 2:
x = x.unsqueeze(0)
attention_mask = torch.ones((x.shape[0], x.shape[1])).to(x.device)
transformer_output = self.transformer(
inputs_embeds=x,
attention_mask=attention_mask,
return_dict=True
)
pooled_output = transformer_output.last_hidden_state[:, 0, :]
return self.classifier(pooled_output)
# Load the model from Huggingface Hub
def load_model():
model_path = hf_hub_download(repo_id="polygraf-ai/vexon-voice-authentication", filename="model.pth")
config_path = hf_hub_download(repo_id="polygraf-ai/vexon-voice-authentication", filename="config.pth")
config = torch.load(config_path)
classifier = PretrainedTransformerClassifier(num_classes=config['num_classes']).to(device)
classifier.load_state_dict(torch.load(model_path, map_location=device))
classifier.eval()
return classifier, config
classifier, model_config = load_model()
def extract_embedding(audio_path):
try:
signal, fs = torchaudio.load(audio_path)
signal = signal.to(device)
embedding = speaker_model.encode_batch(signal)
return embedding.cpu().detach().numpy().flatten()
except Exception as e:
print(f"Error processing {audio_path}: {e}")
return None
def verify_speaker(audio_path1, audio_path2):
emb1 = extract_embedding(audio_path1)
emb2 = extract_embedding(audio_path2)
if emb1 is None or emb2 is None:
return None
tensor1 = torch.tensor(emb1).to(device)
tensor2 = torch.tensor(emb2).to(device)
similarity_score = torch.nn.functional.cosine_similarity(
tensor1, tensor2, dim=0
).cpu().item()
return similarity_score
def process_audio(audio1, audio2):
"""
Process two audio files and return authentication results using the pretrained transformer classifier
"""
if audio1 is None or audio2 is None:
return "Please upload both audio files."
temp_path1 = "temp_audio1.wav"
temp_path2 = "temp_audio2.wav"
try:
torchaudio.save(temp_path1,
torchaudio.load(audio1)[0],
torchaudio.load(audio1)[1])
torchaudio.save(temp_path2,
torchaudio.load(audio2)[0],
torchaudio.load(audio2)[1])
score = verify_speaker(temp_path1, temp_path2)
if score is None:
return "Error processing audio files. Please ensure they are valid audio recordings."
with torch.no_grad():
score_tensor = torch.FloatTensor([[score]]).to(device)
output = classifier(score_tensor)
prediction = torch.argmax(output, dim=1).item()
probabilities = torch.softmax(output, dim=1)[0]
confidence = probabilities[prediction].item()
label = "Original" if prediction == 0 else "Deepfake/Impersonation"
result = f"""
πŸ“Š Authentication Results:
πŸ”Ή Similarity Score: {score:.4f}
πŸ”Ή Classification: {label}
πŸ”Ή Confidence: {confidence:.4f}
{'⚠️ Potential Voice Impersonation Detected!' if prediction > 0 else 'βœ… Authentic Voice Match'}
"""
return result
except Exception as e:
return f"An error occurred: {str(e)}"
finally:
if os.path.exists(temp_path1):
os.remove(temp_path1)
if os.path.exists(temp_path2):
os.remove(temp_path2)
css = """
.gradio-container {
font-family: 'IBM Plex Sans', sans-serif;
}
.gr-button {
color: white;
border-radius: 8px;
background: linear-gradient(to right, #2125ff, #4146ff);
border: none;
cursor: pointer;
}
.gr-button:hover {
background: linear-gradient(to right, #1f23e6, #3b40e6);
}
.footer {
margin-top: 20px;
text-align: center;
border-top: 1px solid #ccc;
padding-top: 10px;
}
"""
demo = gr.Interface(
fn=process_audio,
inputs=[
gr.Audio(label="Reference Voice Recording", type="filepath"),
gr.Audio(label="Voice Recording to Verify", type="filepath")
],
outputs=gr.Textbox(label="Authentication Results"),
title="Vexon Voice Authentication",
description="""
Upload two voice recordings to verify if they are from the same person and detect potential voice impersonation attempts.
The system uses a pretrained transformer model fine-tuned on voice similarity scores to:
1. Calculate a similarity score between the recordings
2. Classify the comparison as Real, DeepFake, or Impersonation
3. Provide a confidence score for the classification
Note: For best results, ensure recordings are clear and contain speech content.
""",
css=css
)
if __name__ == "__main__":
demo.launch()