README.md

---
tags:
- llm
- radio_signal
- signal
- audio
- rf
---
# 📡 RF Signal Classification Model

A Transformer-based model for Radio Frequency (RF) signal classification using spectrogram representations of SDR (Software Defined Radio) recordings.

This model leverages Hugging Face Transformers and the ASTForAudioClassification architecture (Audio Spectrogram Transformer – AST) to classify RF signals directly from audio recordings.

## 🚀 Model Overview

* Architecture: ASTForAudioClassification
* Input: SDR audio recordings (e.g., .wav, .mp3)
* Feature Extraction: AutoFeatureExtractor
* Framework: Hugging Face Transformers
* Task: RF signal classification
The model expects audio sampled at 16 kHz and processes waveform inputs into spectrogram features internally.

## USAGE EXAMPLE

Example signal: ACARS [zip_file](https://huggingface.co/Neo111x/rf_classifier/blob/main/ACARS_sound.mp3.zip)

<audio controls src="https://cdn-uploads.huggingface.co/production/uploads/64d3db80aea0ccb1b4975d95/tFccUxu_EvCkKY2UQTwgB.mpga"></audio>

```python
from transformers import AutoFeatureExtractor, ASTForAudioClassification
from datasets import Dataset, Audio
import torch
import matplotlib.pyplot as plt

# Load feature extractor and model
feature_extractor = AutoFeatureExtractor.from_pretrained("Neo111x/rf_classifier")
model = ASTForAudioClassification.from_pretrained("Neo111x/rf_classifier")

# Use the path to your SDR record file
data = {"audio": ["./ACARS_sound.mp3"]}

# Create dataset and cast to Audio feature (16kHz expected)
dataset = Dataset.from_dict(data).cast_column("audio", Audio(sampling_rate=16000))

# Decode audio
audio_data = dataset[0]["audio"]
audio_array = audio_data["array"]
sampling_rate = audio_data["sampling_rate"]

# Extract features
inputs = feature_extractor(audio_array, sampling_rate=sampling_rate, return_tensors="pt")

# Run inference
with torch.no_grad():
    logits = model(**inputs).logits

predicted_class_ids = torch.argmax(logits, dim=-1).item()
predicted_label = model.config.id2label[predicted_class_ids]
print("Predicted Label:", predicted_label)

# Compute loss (example with target label)
target_label = model.config.id2label[0]
inputs["labels"] = torch.tensor([model.config.label2id[target_label]])
loss = model(**inputs).loss
print("Loss:", round(loss.item(), 2))

# Plot spectrogram (waterfall)
plt.figure(figsize=(12, 6))
plt.specgram(audio_array, NFFT=1024, Fs=sampling_rate, cmap='viridis')
plt.title("SDR Waterfall Plot")
plt.xlabel("Time (s)")
plt.ylabel("Frequency (Hz)")
plt.colorbar(label="Intensity (dB)")
plt.show()
```

## 📊 Input Requirements
* Audio format: .wav, .mp3, or other supported formats
* Sampling rate: 16,000 Hz
* Mono audio preferred
* SDR recordings converted to baseband audio

## 📈 Output
The model outputs:
Predicted label (RF signal class)
Logits for each class
Optional loss value (if labels are provided)
Example output:

```
Predicted Label: ACARS
Loss: 0.12
```


![image](https://cdn-uploads.huggingface.co/production/uploads/64d3db80aea0ccb1b4975d95/vdggy9ssLzgiA9jrGjmgY.png)

## Supported classes

    * 4G LTE Network
    * 5G "New Radio" cellular network - Downlink
    * Aircraft Communications Addressing and Reporting System (ACARS)
    * Amplitude Modulation (AM)
    * Automatic Identification System (AIS)
    * Automatic Link Set-up (ALIS)
    * Automatic Picture Transmission (APT)
    * Bluetooth
    * Differential Global Positioning System (DGPS)
    * Digital Audio Broadcasting Plus (DAB+)
    * Digital Mobile Radio (DMR)
    * Digital Video Broadcasting — Terrestrial (DVB-T)
    * High Frequency Data Link (HFDL)
    * Instrument Landing System
    * M20 Radiosonde
    * Morse Code (CW)
    * Non-Directional Beacon (NDB)
    * Radar altimeter
    * STANAG 5065
    * Secondary surveillance radar (SSR)
    * Single Sideband Voice
    * Tetrapol
    * VHF Data Link - Mode 2 (VDL-M2)
    * VHF Omnidirectional Range (VOR)