app.py

import whisper  # Import whisper from OpenAI
import streamlit as st
import librosa
import numpy as np
import matplotlib.pyplot as plt
from pydub import AudioSegment
from transformers import pipeline
import os

# Load pre-trained sentiment analysis model
sentiment_analyzer = pipeline("sentiment-analysis")

# Streamlit UI
st.title("🎤 Audio Sentiment Analysis")
st.write("Upload multiple MP3 files to analyze sentiment and tone.")

# Upload multiple audio files
uploaded_files = st.file_uploader("Choose MP3 files", type=["mp3"], accept_multiple_files=True)

# Function to process audio, get sentiment, and transcribe to text using Whisper
def analyze_audio(file_path):
    # Convert MP3 to WAV
    audio = AudioSegment.from_mp3(file_path)
    wav_path = file_path.replace(".mp3", ".wav")
    audio.export(wav_path, format="wav")

    # Load audio
    y, sr = librosa.load(wav_path, sr=None)
    
    # Extract MFCCs (Mel-frequency cepstral coefficients)
    mfccs = librosa.feature.mfcc(y=y, sr=sr, n_mfcc=13)
    mfccs_mean = np.mean(mfccs, axis=1)  # Take mean across time axis

    # Perform speech-to-text transcription using Whisper
    model = whisper.load_model("base")  # Load Whisper model (change model size as needed)
    result = model.transcribe(wav_path)
    transcription = result['text']
    
    # Perform sentiment analysis on the transcription
    sentiment_result = sentiment_analyzer(transcription) if transcription else [{"label": "NEGATIVE", "score": 0.5}]
    
    # Remove WAV file after processing
    os.remove(wav_path)
    
    return sentiment_result[0], mfccs_mean, mfccs, y, sr, transcription

# Process and plot if files are uploaded
if uploaded_files:
    # Create a directory to store temporary files
    os.makedirs("temp", exist_ok=True)

    # Prepare a single plot
    fig, ax = plt.subplots(figsize=(10, 6))

    for uploaded_file in uploaded_files:
        # Save the uploaded file
        file_path = f"temp/{uploaded_file.name}"
        with open(file_path, "wb") as f:
            f.write(uploaded_file.getbuffer())

        # Analyze sentiment, extract features, and get transcription
        sentiment, mfccs_mean, mfccs, audio_data, sample_rate, transcription = analyze_audio(file_path)

        # Display sentiment and transcription result
        st.subheader(f"📊 Sentiment Analysis Result for {uploaded_file.name}")
        st.write(f"**Transcription:** {transcription}")
        st.write(f"**Sentiment:** {sentiment['label']}")
        st.write(f"**Confidence:** {sentiment['score']:.2f}")

        # Normalize sentiment score to range from 0 to 1
        sentiment_score = sentiment['score'] if sentiment['label'] == 'POSITIVE' else 1 - sentiment['score']

        # Plotting both curves in a single plot
        ax.plot(np.linspace(0, len(audio_data) / sample_rate, len(audio_data)), [sentiment_score] * len(audio_data), label="Agent's Tone", linestyle='-', color='b')
        
        # Ensure that MFCCs are being averaged across time correctly
        # Take the mean along the time axis (axis=1) to get one value per MFCC coefficient
        mfccs_mean = np.mean(mfccs, axis=1)

        # Plot Patient's Tone Curve (MFCC mean value per coefficient)
        ax.plot(np.linspace(0, len(mfccs_mean), len(mfccs_mean)), mfccs_mean, label="Patient's Tone Curve (MFCCs)", linestyle='--', color='r')

    # Customize plot
    ax.set_xlabel("Time (seconds) / MFCC Coefficients")
    ax.set_ylabel("Score / Mean MFCC Value")
    ax.set_title("Agent Tone & Patient Tone Curve")
    ax.legend()

    # Show plot in Streamlit
    st.pyplot(fig)

    # Clean up temp files
    os.remove(file_path)