SPECTRA - COMINT Signal Analysis Workbench (SIH 2026, PS 26147)

Blind analysis of .wav / .iq intercepts: every parameter is measured from the samples.

Pipeline: burst detection -> baud rate (cyclostationary) -> modulation classification (BPSK / QPSK / 16-QAM / 2-FSK) -> matched filter + timing tracking -> carrier-offset + phase recovery -> soft demapping -> sync-word search (all phase ambiguities / spectrum inversion) -> blind interleaver + FEC detection by trial decoding -> soft Viterbi / Reed-Solomon / concatenated / LDPC decoding -> CRC-checked frame and payload.

Run

pip install -r requirements.txt
python main.py                # GUI  (click Load File, then "Run Full Pipeline")
python test_pipeline.py       # self-test: must print ALL PASS
python demo_signals.py        # regenerate the demo captures

Demo files (ground truth in *.truth.json)

file	modulation	FEC	interleaver	IQ settings
demo.wav	QPSK 2400 baud	Conv K=7	Block 16x16	WAV
demo_bpsk_rs.wav	BPSK 1200	RS(255,223)	none	WAV
demo_qpsk_concat.iq	QPSK 20 kbaud	RS + Conv	Forney depth 12	250000 Hz, cf32
demo_fsk.wav	2-FSK 1200	Conv K=7	Diagonal 8x32	WAV
demo_16qam_uncoded.wav	16-QAM 2400	none	none	WAV

AI Analyst

Set your own key (never commit it): set GEMINI_API_KEY=your_key (or put GEMINI_API_KEY=... in a git-ignored .env). Optional GEMINI_MODEL overrides the model name. Without a key the tab uses an offline rule-based analyst.

Honest limits

Supported demodulators: BPSK, QPSK, 16-QAM, 2-FSK (8-PSK is detected only). FEC: conv K=7 r=1/2 (171,133), RS(255,223) (reedsolo defaults, not the CCSDS dual-basis form), built-in LDPC n=1020. Pseudo-random interleavers need a known seed. Sync library: CCSDS ASM and a few common markers - add yours in the hex box.

BER vs Eb/N0 (Tab 8 and python benchmark.py)

Measured through the real receiver (RRC pulse, 120 Hz carrier offset, timing offset, AWGN -> timing/carrier recovery -> soft demap -> decoder). Curves: uncoded QPSK (with theory overlay), Conv K=7 hard, Conv K=7 soft, LDPC soft. Results are stored in results/ber_vs_snr.png, ber_vs_snr_dark.png (for slides) and ber_vs_snr.csv. Frames with >30 % bit errors are counted as receiver lock failures (listed in the CSV, hollow markers), not as BER. python benchmark.py --quick finishes in seconds; the Tab 8 buttons run the same code.

Web showcase (Vercel)

web/ is a static site (no backend) that displays the receiver's measured results on the five demo captures plus the BER curves. Regenerate its data with python build_site.py; deploy on Vercel with Root Directory = web (framework: Other). The live, interactive tool is the desktop app (python main.py).

For judges - three ways to see the full application

1. Live web version (no install): streamlit_app.py runs the same backend in a browser - upload a .wav/.iq or pick a bundled capture, click Run full pipeline and walk through all 8 tabs. Deployed from this repo via the included Dockerfile (Hugging Face Spaces / Render).
2. Windows app: download SPECTRA-Windows.zip from the Releases page (built by GitHub Actions), unzip, run SPECTRA.exe.
3. From source: pip install -r requirements.txt then python main.py (desktop GUI) or streamlit run streamlit_app.py (web).