NitishBorthakur's picture
Create app.py
f14265e verified
import gradio as gr
import librosa
import numpy as np
import pandas as pd
from sklearn.metrics.pairwise import cosine_similarity
import torch
from speechbrain.inference.speaker import EncoderClassifier
from sklearn.decomposition import PCA
from sklearn.manifold import TSNE
import plotly.graph_objects as go
from sklearn.preprocessing import normalize
import os
from cryptography.fernet import Fernet
import pickle
# --- Configuration using Environment Variables ---
encrypted_file_path = os.environ.get("SPEAKER_EMBEDDINGS_FILE")
metadata_file = os.environ.get("METADATA_FILE")
visualization_method = os.environ.get("VISUALIZATION_METHOD", "pca")
max_length = 5 * 16000
num_closest_speakers = 20
pca_dim = 50
# --- Check for Missing Environment Variables ---
if not encrypted_file_path:
raise ValueError("SPEAKER_EMBEDDINGS_FILE environment variable is not set.")
if not metadata_file:
raise ValueError("METADATA_FILE environment variable is not set.")
# --- Check for valid visualization method ---
if visualization_method not in ["pca", "tsne"]:
raise ValueError("Invalid VISUALIZATION_METHOD. Choose 'pca' or 'tsne'.")
# --- Debugging: Check Environment Variables ---
print(f"DECRYPTION_KEY: {os.getenv('DECRYPTION_KEY')}")
print(f"SPEAKER_EMBEDDINGS_FILE: {os.getenv('SPEAKER_EMBEDDINGS_FILE')}")
if os.getenv('SPEAKER_EMBEDDINGS_FILE'):
print(
f"Encrypted file path exists: {os.path.exists(os.getenv('SPEAKER_EMBEDDINGS_FILE'))}"
)
else:
print(
"Encrypted file path does not exist: SPEAKER_EMBEDDINGS_FILE environment variable not set or file not found."
)
# --- Decryption ---
key = os.getenv("DECRYPTION_KEY")
if not key:
raise ValueError(
"Decryption key is missing. Ensure DECRYPTION_KEY is set in the environment variables."
)
fernet = Fernet(key.encode("utf-8"))
# --- Sample Audio Files ---
sample_audio_dir = "sample_audio"
sample_audio_files = [
"Bob_Barker.mp3",
"Howie_Mandel.m4a",
"Katherine_Jenkins.mp3",
]
# --- Load Embeddings and Metadata ---
try:
with open(encrypted_file_path, "rb") as encrypted_file:
encrypted_data = encrypted_file.read()
decrypted_data_bytes = fernet.decrypt(encrypted_data)
# Deserialize using pickle.loads()
speaker_embeddings = pickle.loads(decrypted_data_bytes)
print("Speaker embeddings loaded successfully!")
except FileNotFoundError:
raise FileNotFoundError(
f"Could not find encrypted embeddings file at: {encrypted_file_path}"
)
except Exception as e:
raise Exception(f"Error during decryption or loading embeddings: {e}")
df = pd.read_csv(metadata_file, delimiter="\t")
# --- Convert Embeddings to NumPy Arrays ---
for spk_id, embeddings in speaker_embeddings.items():
speaker_embeddings[spk_id] = [np.array(embedding) for embedding in embeddings]
# --- Speaker ID to Name Mapping ---
speaker_id_to_name = dict(zip(df["VoxCeleb1 ID"], df["VGGFace1 ID"]))
# --- Load SpeechBrain Classifier ---
classifier = EncoderClassifier.from_hparams(
source="speechbrain/spkrec-xvect-voxceleb",
savedir="pretrained_models/spkrec-xvect-voxceleb",
)
# --- Function to Calculate Average Embedding (Centroid) ---
def calculate_average_embedding(embeddings):
avg_embedding = np.mean(embeddings, axis=0)
return normalize(avg_embedding.reshape(1, -1)).flatten()
# --- Precompute Speaker Centroids ---
speaker_centroids = {
spk_id: calculate_average_embedding(embeddings)
for spk_id, embeddings in speaker_embeddings.items()
}
# --- Function to Prepare Data for Visualization ---
def prepare_data_for_visualization(speaker_centroids, closest_speaker_ids):
all_embeddings = [
centroid
for speaker_id, centroid in speaker_centroids.items()
if speaker_id in closest_speaker_ids
]
all_speaker_ids = [
speaker_id
for speaker_id in speaker_centroids
if speaker_id in closest_speaker_ids
]
return np.array(all_embeddings), np.array(all_speaker_ids)
# --- Function to Reduce Dimensionality ---
def reduce_dimensionality(all_embeddings, method="tsne", perplexity=5, pca_dim=50):
if method == "pca":
reducer = PCA(n_components=2)
elif method == "tsne":
pca_reducer = PCA(n_components=pca_dim)
all_embeddings = pca_reducer.fit_transform(all_embeddings)
reducer = TSNE(n_components=2, random_state=42, perplexity=perplexity)
else:
raise ValueError("Invalid method. Choose 'pca' or 'tsne'.")
reduced_embeddings = reducer.fit_transform(all_embeddings)
return reducer, reduced_embeddings
# --- Function to Get Speaker Name from ID ---
def get_speaker_name(speaker_id):
return speaker_id_to_name.get(speaker_id, f"Unknown ({speaker_id})")
# --- Function to Generate Visualization ---
def generate_visualization(
pca_reducer,
reduced_embeddings,
all_speaker_ids,
new_embedding,
predicted_speaker_id,
visualization_method,
perplexity,
pca_dim,
):
if visualization_method == "pca":
new_embedding_reduced = pca_reducer.transform(new_embedding.reshape(1, -1))
elif visualization_method == "tsne":
combined_embeddings = np.vstack(
[reduced_embeddings, new_embedding.reshape(1, -1)]
)
reducer = TSNE(n_components=2, random_state=42, perplexity=perplexity)
combined_reduced = reducer.fit_transform(combined_embeddings)
reduced_embeddings = combined_reduced[:-1]
new_embedding_reduced = combined_reduced[-1].reshape(1, -1)
else:
raise ValueError("Invalid visualization method.")
fig = go.Figure()
fig.add_trace(
go.Scatter(
x=reduced_embeddings[:, 0],
y=reduced_embeddings[:, 1],
mode="markers",
marker=dict(color="blue", size=8, opacity=0.5),
text=[get_speaker_name(speaker_id) for speaker_id in all_speaker_ids],
name="Other Speakers",
)
)
if predicted_speaker_id in all_speaker_ids:
predicted_speaker_index = list(all_speaker_ids).index(predicted_speaker_id)
fig.add_trace(
go.Scatter(
x=[reduced_embeddings[predicted_speaker_index, 0]],
y=[reduced_embeddings[predicted_speaker_index, 1]],
mode="markers",
marker=dict(
color="green",
size=10,
symbol="circle",
line=dict(color="black", width=2),
),
name=get_speaker_name(predicted_speaker_id),
text=[get_speaker_name(predicted_speaker_id)],
)
)
fig.add_trace(
go.Scatter(
x=new_embedding_reduced[:, 0],
y=new_embedding_reduced[:, 1],
mode="markers",
marker=dict(color="red", size=12, symbol="star"),
name="New Voice",
text=["New Voice"],
)
)
fig.update_layout(
title=f"Dimensionality Reduction of Speaker Embeddings using {visualization_method.upper()}",
xaxis_title="Component 1",
yaxis_title="Component 2",
legend=dict(x=0, y=1, traceorder="normal", orientation="h"),
hovermode="closest",
)
return fig
# --- Main Function ---
def identify_voice_and_visualize_with_averaging(audio_file, perplexity=5):
try:
if isinstance(audio_file, str):
signal, fs = librosa.load(audio_file, sr=16000)
elif isinstance(audio_file, np.ndarray):
signal = audio_file
fs = 16000
else:
raise ValueError(
"Invalid audio input. Must be a file path or a NumPy array."
)
signal_tensor = torch.tensor(signal, dtype=torch.float32).unsqueeze(0)
signal_tensor = torch.nn.functional.pad(
signal_tensor, (0, max_length - signal_tensor.shape[1])
)
user_embedding = classifier.encode_batch(signal_tensor).cpu().detach().numpy()
user_embedding = normalize(
user_embedding.squeeze(axis=(0, 1)).reshape(1, -1)
).flatten()
similarity_scores = {
spk_id: cosine_similarity(
user_embedding.reshape(1, -1), centroid.reshape(1, -1)
)[0][0]
for spk_id, centroid in speaker_centroids.items()
}
closest_speaker_ids = sorted(
similarity_scores, key=similarity_scores.get, reverse=True
)[:num_closest_speakers]
predicted_speaker_id = closest_speaker_ids[0]
highest_similarity = similarity_scores[predicted_speaker_id]
all_embeddings, all_speaker_ids = prepare_data_for_visualization(
speaker_centroids, closest_speaker_ids
)
reducer, reduced_embeddings = reduce_dimensionality(
all_embeddings,
method=visualization_method,
perplexity=perplexity,
pca_dim=pca_dim,
)
predicted_speaker_name = get_speaker_name(predicted_speaker_id)
similarity_percentage = round(highest_similarity * 100, 2) # Rounded here
visualization = generate_visualization(
reducer,
reduced_embeddings,
all_speaker_ids,
user_embedding,
predicted_speaker_id,
visualization_method,
perplexity,
pca_dim,
)
result_text = (
f"The voice resembles speaker: {predicted_speaker_name} "
f"with a similarity of {similarity_percentage:.2f}%" # Display rounded value
)
return result_text, visualization
except Exception as e:
return f"Error during processing: {e}", None
# --- Gradio Interface ---
# Create a directory for caching examples if it doesn't exist
cache_dir = "examples_cache"
if not os.path.exists(cache_dir):
os.makedirs(cache_dir)
# Define the Gradio interface
iface = gr.Interface(
fn=identify_voice_and_visualize_with_averaging,
inputs=gr.Audio(type="filepath", label="Input Audio"),
outputs=["text", gr.Plot()],
title="Discover Your Celebrity Voice Twin!",
description="Record your voice or upload an audio file, and see your celebrity match! Not ready to record? Try our sample voices to see how it works!",
cache_examples=False,
examples_per_page=3,
examples=[
[os.path.join(sample_audio_dir, sample_audio_files[0])],
[os.path.join(sample_audio_dir, sample_audio_files[1])],
[os.path.join(sample_audio_dir, sample_audio_files[2])],
],
)
# Launch the interface
iface.launch(debug=True, share=True)