--- license: mit tags: - signal-processing - modulation-classification - radioml - pytorch - dsp library_name: pytorch pipeline_tag: audio-classification spaces: - alirezaaminzadeh/radio-modulation-classifier model-index: - name: alirezaaminzadeh/radio-modulation-classifier results: - task: type: audio-classification name: Modulation Classification dataset: name: RadioML 2018.01A type: radioml-2018 metrics: - type: accuracy value: 0.7903 name: test_accuracy_all_snr - type: accuracy value: 0.9365 name: test_accuracy_snr_gte_0db - type: f1 name: macro_f1_all_snr value: 0.7878 - type: f1 name: macro_f1_snr_gte_0db value: 0.9390 --- # SignalIQ Modulation Classifier Convolutional LSTM Deep Network (CLDNN) for automatic radio modulation recognition from I/Q samples. ## Model Description - **Architecture:** Convolutional LSTM Deep Network (CLDNN) - **Input:** 2 × 1024 float tensor (In-Phase and Quadrature channels) - **Output:** 11 modulation classes (RadioML 2018.01A) ### Supported Modulations AM-DSB, AM-SSB, WBFM, QPSK, QAM16, QAM64, 8PSK, BPSK, CPFSK, GFSK, PAM4 ## Intended Uses - RF spectrum monitoring and signal intelligence - IoT device RF fingerprinting and security - Educational demonstrations of DSP + deep learning ## Out-of-Scope Uses - Unauthorized signal interception or surveillance - Real-time SDR deployment without domain validation - Modulation types not in the RadioML 2018.01A taxonomy ## Training Data Training uses the [RadioML 2018.01A](https://www.deepsig.ai/datasets) taxonomy (11 classes, SNR -20 to +30 dB). When the full pickle is unavailable locally, a **synthetic I/Q generator** matching RadioML modulation characteristics is used (stratified 70/15/15 split). Re-train with the downloaded dataset for full-benchmark scores. ## Evaluation Results | Metric | Value | |--------|-------| | Test Accuracy (all SNR) | 0.7903 | | Macro F1 (all SNR) | 0.7878 | | Accuracy (SNR ≥ 0 dB) | 0.9365 | | Macro F1 (SNR ≥ 0 dB) | 0.9390 | ## Limitations - Performance degrades significantly below -10 dB SNR - QAM16/QAM64 and QPSK are harder classes (lower F1 at low SNR) - Trained on synthetic channel models when RadioML pickle is unavailable; real-world performance may vary - Fixed input length of 1024 samples ## Ethical Considerations This model has dual-use potential in RF classification. It must not be used for unauthorized interception of communications. ## License MIT License — see LICENSE file. ## Citation ```bibtex @misc{signaliq-modulation-classifier-100, title={SignalIQ Modulation Classifier}, author={Aria AI Engineering}, year={2026}, publisher={Hugging Face}, howpublished={\url{https://huggingface.co/alirezaaminzadeh/radio-modulation-classifier}} } ``` ## Related Resources - **Space:** [Interactive Demo](https://huggingface.co/spaces/alirezaaminzadeh/radio-modulation-classifier) - **Dataset:** [Sample I/Q Dataset](https://huggingface.co/datasets/alirezaaminzadeh/radioml-2018-samples)