🛡️ Forensic Audio Classifier Tool

Hybrid AM + LM + Classifier Based Crime Speech Detection System
Optimized for Tripura Bengali Dialect

👤 Developer Information

Developer: Arnab Das
Institute: NFSU Tripura Campus
Year: 2025

📌 Overview

The Forensic Audio Classifier Tool is a machine-learning based forensic audio analysis system designed to transcribe, analyze, and classify speech content in Tripura’s Bengali dialect, influenced by:

• Comilla
• Dhaka
• Noakhali
• Sylhet
• Bengali (West Bengal)

The tool automatically categorizes audio evidence into:

• 🚨 Flagged — Crime-related audio
• 🧐 Review — Ambiguous or partially suspicious
• ✅ Safe — Harmless speech

---

⚙️ System Capabilities

This tool integrates multiple forensic speech-processing components:

• Acoustic Model: Wav2Vec2-BERT (Bangla ASR)
• Language Model: 3-gram KenLM (Tripura Bengali Corpus)
• Classifier Model: Fine-tuned Crime vs. Non-Crime Transformer
• Keyword Engine: Weighted, fuzzy matching & severity scoring
• Decision Logic: Hybrid probability + rule-based detection
• Pre-processing: Silence trimming & audio normalization
• Output Formats: HTML & CSV forensic reports
• Audio Segregation: Flagged / Review / Safe evidence folders

---

🧠 System Architecture

The complete forensic processing workflow consists of:

• Audio Input
• Acoustic Model (Wav2Vec2-BERT)
• Language Model (3-gram KenLM)
• Transcription Generation
• Keyword Detection Engine
• Crime Classification Model
• Hybrid Decision Logic
• Safe / Review / Flagged Categorization
• HTML & CSV Report Generation

---

📦 Components

🔊 Acoustic Model (Wav2Vec2-BERT)

• Source: https://huggingface.co/sazzadul/Shrutimala_Bangla_ASR
• License: CC-BY-SA 4.0
• Creator: Md Sazzadul Islam Ridoy
• Integrated by: Arnab Das

---

📝 3-Gram Language Model (KenLM)

• Developed by: Arnab Das
• Corpus: Custom Tripura-Bengali dataset
• Purpose: Improve decoding accuracy & dialect handling

---

🔍 Crime Classifier Model

• Model Type: Transformer-based text classifier
• Labels: Crime / Non-Crime
• Fine-tuned by: Arnab Das
• Base Model: Deleted by original author (self-maintained)

---

📁 Folder Structure

Forensic-Audio-Classifier-Tool/
│
├── tool.py
├── config.json
├── requirements.txt
├── README.md
│
├── models/
│   ├── language_model/
│   │     ├── lm.arpa
│   │     └── unigrams.txt
│   └── classifier_model/
│
├── keywords/
│   ├── critical.csv
│   ├── high.csv
│   ├── medium.csv
│   ├── low.csv
│   └── context_tokens.csv
│
└── sample_outputs/
      ├── report_example.html
      └── report_example.csv

---

🚀 How to Run

Follow the steps below to execute the forensic analysis tool:

---

1️⃣ Install Dependencies

pip install -r requirements.txt

---

2️⃣ Place Your Audio Files

mkdir audios
# place your .wav/.mp3/.opus files in this folder

---

3️⃣ Run the Forensic Tool

python tool.py

---

4️⃣ Generated Outputs

📄 Reports

crime_detection_report_<timestamp>.html
crime_detection_report_<timestamp>.csv

📂 Segregated Audio Folders

flagged_audio/
review_audio/
safe_audio/

---

📚 Citation (APA)

Arnab Das, Forensic Audio Classifier Tool (v0.9): Hybrid AM + LM + Classifier System for Crime-Related Speech Detection in Tripura Bengali, Technical Report, National Forensic Sciences University (NFSU), Tripura Campus, 2025.

---

👨‍💻 Author

• Name: Arnab Das
• Field: Forensic Audio Research & Development
• Institute: NFSU Tripura Campus (2025)