Upload 6 files

app.py

@@ -0,0 +1,44 @@
+import streamlit as st
+import librosa
+import soundfile as sf
+import tempfile
+import os
+from utils.noise_removal import remove_noise
+from utils.vad_segmentation import detect_speech_segments
+from utils.speaker_diarization import diarize_speakers
+from utils.noise_classification import classify_noise
+
+st.set_page_config(page_title="Audio Analyzer", layout="wide")
+st.title(" Audio Analysis Pipeline")
+
+uploaded_file = st.file_uploader("Upload an audio file", type=["wav", "mp3"])
+
+if uploaded_file:
+    with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as tmp:
+        tmp.write(uploaded_file.read())
+        tmp_path = tmp.name
+
+    st.audio(tmp_path, format='audio/wav')
+
+    st.subheader("1️⃣ Noise Removal")
+    denoised_path = tmp_path.replace(".wav", "_denoised.wav")
+    remove_noise(tmp_path, denoised_path)
+    st.audio(denoised_path, format="audio/wav")
+
+    st.subheader("2️⃣ Speech Segmentation")
+    speech_segments = detect_speech_segments(denoised_path)
+    st.write(f"Detected {len(speech_segments)} speech segments.")
+    for i, (start, end) in enumerate(speech_segments[:5]):
+        st.write(f"Segment {i+1}: {start:.2f}s to {end:.2f}s")
+
+    st.subheader("3️⃣ Speaker Diarization")
+    diarization = diarize_speakers(denoised_path)
+    st.text("Speakers detected:")
+    for turn, _, speaker in diarization.itertracks(yield_label=True):
+        st.write(f"{turn.start:.2f}s - {turn.end:.2f}s: {speaker}")
+
+    st.subheader("4️⃣ Noise Classification")
+    noise_predictions = classify_noise(denoised_path)
+    st.write("Top predicted noise classes:")
+    for cls, prob in noise_predictions:
+        st.write(f"{cls}: {prob:.2f}")

requirements.txt

@@ -0,0 +1,7 @@
+streamlit
+soundfile
+librosa
+speechbrain
+pyannote.audio
+torchaudio
+scikit-learn

utils/noise_classification.py

@@ -0,0 +1,15 @@
+import numpy as np
+import librosa
+import joblib
+
+# Load your trained model + label encoder
+clf = joblib.load("models/noise_classifier.pkl")
+label_encoder = joblib.load("models/label_encoder.pkl")
+
+def classify_noise(audio_path):
+    y, sr = librosa.load(audio_path, sr=None)
+    mfcc = librosa.feature.mfcc(y=y, sr=sr, n_mfcc=13)
+    feature = np.mean(mfcc.T, axis=0).reshape(1, -1)
+    probs = clf.predict_proba(feature)[0]
+    top_idx = np.argsort(probs)[::-1][:5]
+    return [(label_encoder.inverse_transform([i])[0], probs[i]) for i in top_idx]

utils/noise_removal.py

@@ -0,0 +1,8 @@
+from speechbrain.pretrained import SpectralMaskEnhancement
+import torchaudio
+
+model = SpectralMaskEnhancement.from_hparams(source="speechbrain/metricgan-plus-voicebank")
+
+def remove_noise(input_path, output_path):
+    enhanced = model.enhance_file(input_path)
+    torchaudio.save(output_path, enhanced[0], enhanced[1])

utils/speaker_diarization.py

@@ -0,0 +1,6 @@
+from pyannote.audio import Pipeline
+
+diarization_pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization")
+
+def diarize_speakers(audio_path):
+    return diarization_pipeline(audio_path)

utils/vad_segmentation.py

@@ -0,0 +1,8 @@
+import torchaudio
+from pyannote.audio import Pipeline
+
+pipeline = Pipeline.from_pretrained("pyannote/voice-activity-detection")
+
+def detect_speech_segments(audio_path):
+    vad_result = pipeline(audio_path)
+    return [(segment.start, segment.end) for segment in vad_result.get_timeline().support()]