⚡ Trim audio to 30s for stable analysis on Hugging Face

app.py

@@ -29,32 +29,41 @@ def analyze_emotion(audio_path):
     extractor, model = load_model()
     waveform, sr = torchaudio.load(audio_path)
 
-    if sr != 16000:
-        waveform = torchaudio.transforms.Resample(sr, 16000)(waveform)
+    # 💡 Trim audio to 30 seconds max to avoid slowdowns
+    max_duration_sec = 30
+    max_samples = sr * max_duration_sec
+    if waveform.size(1) > max_samples:
+        waveform = waveform[:, :max_samples]
 
+    duration_sec = waveform.size(1) / sr
+
+    # Run model
     inputs = extractor(waveform[0].numpy(), sampling_rate=16000, return_tensors="pt")
     with torch.no_grad():
         logits = model(**inputs).logits[0]
 
     emotion, scores = get_emotion_label(logits)
-    return emotion.capitalize(), scores
+    return emotion.capitalize(), scores, duration_sec
 
-# UI
+# Streamlit UI
 st.set_page_config(page_title="🎧 Audio Emotion Detector", layout="centered")
 st.title("🎧 Audio Emotion Analysis (Wav2Vec2)")
 
 uploaded_file = st.file_uploader("Upload an MP3 or WAV audio file", type=["mp3", "wav"])
 
 if uploaded_file:
-    st.audio(uploaded_file)
+    st.audio(uploaded_file, format='audio/wav')
     with st.spinner("Analyzing emotion..."):
         wav_path = convert_to_wav(uploaded_file)
-        emotion, scores = analyze_emotion(wav_path)
+        emotion, scores, duration_sec = analyze_emotion(wav_path)
+
+        st.subheader("⏱ Audio Info:")
+        st.write(f"Duration analyzed: **{duration_sec:.2f} seconds**")
 
-        st.subheader("Detected Emotion:")
-        st.markdown(f"🧠 **{emotion}**")
+        st.subheader("🧠 Detected Emotion:")
+        st.markdown(f"**{emotion}**")
 
-        st.subheader("Confidence Scores:")
+        st.subheader("🎯 Confidence Scores:")
         emotions = ["angry", "happy", "neutral", "sad"]
         for i, label in enumerate(emotions):
             st.write(f"- **{label.capitalize()}**: {scores[i]*100:.2f}%")