Update README.md

README.md

@@ -9,137 +9,26 @@ metrics:
 - accuracy
 - f1
 - roc_auc
+- recall
+- precision
 base_model:
 - facebook/wav2vec2-base-960h
 - openai/whisper-large-v3-turbo
+- MelodyMachine/Deepfake-audio-detection-V2
 pipeline_tag: audio-classification
 ---
-import numpy as np
-import librosa
-import tensorflow as tf
-import streamlit as st
-import sounddevice as sd
-import wave
-import os
-
-# Constants
-window_length = 0.02  # 20ms window length
-hop_length = 0.0025  # 2.5ms hop length
-sample_rate = 22050  # Standard audio sample rate
-n_mels = 128  # Number of mel filter banks
-threshold_zcr = 0.1  # Adjust this threshold to detect breath based on ZCR
-threshold_rmse = 0.1  # Adjust this threshold to detect breath based on RMSE
-max_len = 500  # Fix length for feature extraction
-
-# Load TFLite model
-interpreter = tf.lite.Interpreter(model_path="model_breath_logspec_mfcc_cnn.tflite")
-interpreter.allocate_tensors()
-
-# Get input and output details
-input_details = interpreter.get_input_details()
-output_details = interpreter.get_output_details()
-
-# Function to extract breath features
-def extract_breath_features(y, sr):
-    frame_length = int(window_length * sr)
-    hop_length_samples = int(hop_length * sr)
-    
-    zcr = librosa.feature.zero_crossing_rate(y=y, frame_length=frame_length, hop_length=hop_length_samples)
-    rmse = librosa.feature.rms(y=y, frame_length=frame_length, hop_length=hop_length_samples)
-    
-    zcr = zcr.T.flatten()
-    rmse = rmse.T.flatten()
-    
-    breaths = (zcr > threshold_zcr) & (rmse > threshold_rmse)
-    breath_feature = np.where(breaths, 1, 0)
-    
-    return breath_feature
-
-# Feature extraction
-def extract_features(file_path):
-    try:
-        y, sr = librosa.load(file_path, sr=None)
-        mfcc = librosa.feature.mfcc(y=y, sr=sr, n_mfcc=13)
-        logspec = librosa.amplitude_to_db(librosa.feature.melspectrogram(y=y, sr=sr, n_mels=n_mels))
-        breath_feature = extract_breath_features(y, sr)
-        
-        # Fix lengths
-        mfcc = librosa.util.fix_length(mfcc, size=max_len, axis=1)
-        logspec = librosa.util.fix_length(logspec, size=max_len, axis=1)
-        breath_feature = librosa.util.fix_length(breath_feature, size=max_len)
-        
-        return np.vstack((mfcc, logspec, breath_feature))
-    except Exception as e:
-        st.error(f"Error processing audio: {e}")
-        return None
-
-# Prepare input for model
-def prepare_single_data(features):
-    features = librosa.util.fix_length(features, size=max_len, axis=1)
-    features = features[np.newaxis, ..., np.newaxis]  # Add batch and channel dimensions
-    return features.astype(np.float32)  # Convert to FLOAT32
-
-# Predict audio class
-def predict_audio(file_path):
-    features = extract_features(file_path)
-    if features is not None:
-        prepared_features = prepare_single_data(features)
-        interpreter.set_tensor(input_details[0]['index'], prepared_features)
-        interpreter.invoke()
-        prediction = interpreter.get_tensor(output_details[0]['index'])
-        predicted_class = np.argmax(prediction, axis=1)
-        predicted_prob = prediction[0]
-        return predicted_class[0], predicted_prob
-    return None, None
-
-# Record audio function
-def record_audio(duration=5, samplerate=22050):
-    st.info(f"Recording for {duration} seconds...")
-    audio_data = sd.rec(int(duration * samplerate), samplerate=samplerate, channels=1, dtype=np.int16)
-    sd.wait()
-    st.success("Recording Complete!")
-    return audio_data, samplerate
-
-# Save recorded audio as .wav
-def save_wav(file_path, audio_data, samplerate):
-    with wave.open(file_path, 'wb') as wf:
-        wf.setnchannels(1)
-        wf.setsampwidth(2)
-        wf.setframerate(samplerate)
-        wf.writeframes(audio_data.tobytes())
-
-# Streamlit UI
-st.title('🎙️ Audio Deepfake Detection')
-st.write('Upload or record an audio file to classify it as real or fake.')
-
-# File uploader
-uploaded_file = st.file_uploader('Upload an audio file', type=['wav', 'mp3'])
-recorded_file_path = "recorded_audio.wav"
-
-# Record audio button
-if st.button("Record Live Audio"):
-    duration = st.slider("⏳ Set Duration (seconds)", 1, 10, 5)
-    audio_data, samplerate = record_audio(duration)
-    save_wav(recorded_file_path, audio_data, samplerate)
-    st.audio(recorded_file_path, format="audio/wav")
-
-# Process uploaded or recorded audio
-if uploaded_file is not None:
-    with open("uploaded_audio.wav", 'wb') as f:
-        f.write(uploaded_file.getbuffer())
-    file_path = "uploaded_audio.wav"
-    st.audio(file_path, format="audio/wav")
-elif os.path.exists(recorded_file_path):
-    file_path = recorded_file_path
-else:
-    file_path = None
-
-# Run prediction
-if file_path:
-    prediction, probability = predict_audio(file_path)
-    if prediction is not None:
-        st.write(f'**Predicted Class:** {prediction}')
-        st.write(f'**Probability of being Real:** {probability[0]*100:.2f}%')
-        st.write(f'**Probability of being Fake:** {probability[1]*100:.2f}%')
-    else:
-        st.error("Failed to process the audio file.")
+---
+datasets:
+- DynamicSuperb/SpeechDetection_LJSpeech
+- DynamicSuperb/AudioDeepFakeDetection_LJSpeech_WaveFake_MUSDB18HQ
+- DynamicSuperb/SceneFakeDetection_SceneFake_ASPIRE
+language:
+- en
+metrics:
+- accuracy
+- f1
+- roc_auc
+base_model:
+- facebook/wav2vec2-base-960h
+- openai/whisper-large-v3-turbo
+pipeline_tag: audio-classification