Voice Sentiment Analysis System

Project Description

This project is an automated solution for analyzing customer satisfaction from voice calls. Built using state-of-the-art machine learning models, it combines Wav2Vec 2.0 for speech-to-text transcription with BERT for sentiment analysis to provide real-time feedback into customer emotions and satisfaction levels.

Key Features

• Automatic Speech Recognition: Convert voice calls to text using Wav2Vec 2.0
• Sentiment Analysis: Analyze emotional tone using multilingual BERT
• Customer Satisfaction Classification: Categorize calls as Satisfied, Dissatisfied, or Neutral
• Batch Processing: Handle multiple audio files simultaneously
• Web Interface: User-friendly Gradio interface for easy interaction
• CSV Export: Detailed results export for further analysis and reporting

Live Deployments

REST API (Render)

Production API Endpoint: https://voice-sentiment-analysis-0eh4.onrender.com

Live Demo on Hugging Face powered By Gradio

Interactive Demo: https://huggingface.co/spaces/martinoyovo/voice-sentiment-analysis

Project Structure

voice-sentiment-project/
├── requirements.txt           # Dependencies
├── voice_sentiment.py         # Core analyzer class
├── api.py                     # REST API Server
├── app.py                     # Gradio web interface  
├── main.py                    # CLI interface
├── utils.py                   # Utility functions and CSS styling
├── audios/                     # Your audio files
│   ├── call1.wav
│   ├── call2.mp3
│   └── ...
└── analysis_results.csv       # Generated results

Quick Installation

pip install -r requirements.txt

Usage

1. Web Interface (Recommended)

python app.py

Opens a web browser interface at http://localhost:7860

2. Command Line Interface

python main.py

3. Direct Code Usage

from voice_sentiment import VoiceSentimentAnalyzer

# Initialize
analyzer = VoiceSentimentAnalyzer()

# Analyze one call
result = analyzer.analyze_call("call1.wav")
print(result)

# Analyze multiple calls
results = analyzer.analyze_batch("audios/")

Example Output

{
    'file': 'call1.wav',
    'transcription': 'Hello I am very satisfied with your service',
    'sentiment': 'POSITIVE', 
    'score': 0.89,
    'satisfaction': 'Satisfied'
}

Simple Workflow

Audio File → Transcription (Wav2Vec2) → Sentiment (BERT) → Classification

Perfect for analyzing customer call sentiment quickly and easily!

Main Use Cases

• Customer Service & Call Centers
• Business Intelligence & Analytics
• Financial Services
• Sales & E-commerce
• Healthcare & Telemedicine

Models Used

This project uses pre-trained models hosted on Hugging Face Hub:

Speech Recognition Model

Wav2Vec 2.0 - English

• Model: facebook/wav2vec2-large-960h-lv60-self
• Link: https://huggingface.co/facebook/wav2vec2-large-960h-lv60-self
• Description: Large Wav2Vec 2.0 model trained on 960 hours of English LibriSpeech data
• Use: Audio-to-text transcription

Sentiment Analysis Model

BERT - Multilingual Sentiment

• Model: nlptown/bert-base-multilingual-uncased-sentiment
• Link: https://huggingface.co/nlptown/bert-base-multilingual-uncased-sentiment
• Description: Multilingual BERT model fine-tuned for sentiment analysis (1-5 stars)
• Use: Text sentiment classification

Language Support

Current Model: English Only

This system is currently configured with an English-only Wav2Vec 2.0 model (facebook/wav2vec2-large-960h-lv60-self) for optimal English speech recognition performance.

For Other Languages

To use this system with other languages, you need to change the Wav2Vec 2.0 model in voice_sentiment.py.

Supported Audio Formats

Fully Supported

• WAV (.wav) - Recommended for best quality
• MP3 (.mp3) - Most common format
• M4A (.m4a) - Apple audio format

Audio Specifications

• Sample Rate: Automatically converted to 16kHz
• Channels: Mono or Stereo (converted to mono)
• Duration: 5 seconds to 10 minutes (optimal: 30 seconds - 2 minutes)
• Quality: Clear speech, minimal background noise recommended

Not Supported

• Video files (MP4, AVI, MOV, etc.)
• Other audio formats (FLAC, OGG, etc.) - may work but not guaranteed
• Extremely low quality or heavily distorted audio
• Files with encryption or DRM protection

Audio Quality Tips

• Use WAV format for highest accuracy
• Ensure clear speech recording
• Minimize background noise
• Optimal recording: 16kHz, 16-bit, mono
• Test with short samples first

CSV Output & Results

Automatic CSV Generation

When using batch analysis (multiple files), the system automatically generates a detailed CSV file with all results.

File: analysis_results.csv

Location: Same folder as the project

CSV Contents

File,Transcription,Sentiment,Score,Satisfaction
call1.wav,"Hello I am very satisfied with your service",POSITIVE,0.89,Satisfied
call2.wav,"This is unacceptable I want a refund",NEGATIVE,0.92,Dissatisfied
call3.wav,"Can you tell me about your pricing",NEUTRAL,0.65,Neutral