📻 Tactical Radio Intercept Analyzer

Overview

This repository contains an R&D project focused on the automated classification of tactical radio intercepts. The system utilizes a multi-modal AI pipeline to transcribe noisy, real-world radio communications and classify them into structured data using Large Language Models (LLMs).

Developed as a complex academic research project, this tool demonstrates the practical application of NLP and Audio Processing in military/tactical contexts.

Architecture & Pipeline

The system operates in two main stages:

1. Automatic Speech Recognition (ASR): Uses OpenAI's whisper-large-v3-turbo (via Hugging Face transformers) to transcribe Ukrainian audio, specifically optimized to handle severe static, background noise, and clipping typical of tactical comms.
2. LLM Text Classification: Employs Google's gemini-2.5-flash via LangChain. A strict system prompt and Pydantic output parsers force the model to categorize the messy transcription into predefined tactical clusters (e.g., Reconnaissance, Medevac, Artillery, Logistics) and extract entities like coordinates and callsigns into a clean JSON format.

Features

• End-to-End Processing: From raw .wav/.ogg audio straight to a structured JSON intelligence report.
• Resilient ASR: Capable of understanding heavily distorted speech (simulated radio chatter).
• Hallucination Control: Zero-temperature LLM generation with strict schema enforcement.
• Interactive UI: Built-in Gradio web interface for live audio recording and file uploads.
• DSP Dataset Generator: Includes a custom Digital Signal Processing script (batch_radio_fx.py) to apply bandpass filters, overdrive, and white noise to clean audio, simulating gritty tactical environments.

Quick Start (Local Setup)

Prerequisites

• Python 3.10+
• FFmpeg installed on your system.
• Google Gemini API Key.

Installation

1. Clone the repository:

   git clone https://github.com AlexanderSychev2005/radio-analysis.git
   cd radio-analysis

2. Install dependencies:

   pip install -r requirements.txt

3. Set your API Key:

   export GOOGLE_API_KEY="your_api_key_here"

Running the app

Launch the interactive web interface:

python app.py

Open http://127.0.0.1:7860 in your browser.

Dataset Generation (Radio FX)

To test the model, you can generate your own tactical intercepts. Place clean .wav or .ogg voice recordings in the test_audios/ folder and run:

python radio_fx.py

This script applies mathematical filters (cutting frequencies below 300Hz and above 3000Hz, applying clipping, and adding white noise) and outputs the results to the radio_audios/ folder.

Model Metrics & Evaluation

The ASR component is evaluated based on the Word Error Rate (WER). While the model may produce slight transcription errors due to induced radio static, the subsequent LLM's self-attention mechanism effectively contextualizes and corrects these discrepancies during the classification stage.

---

Developed for research and educational purposes.