Automatic modulation classification (AMC) allows the receiving system to determine the modulation scheme of an incoming signal without requiring prior information about the transmitter's configuration. In hybrid convolutional–recurrent models operating on time–frequency representations, flattening convolutional features into a single sequence step eliminates the explicit sequence of time-ordered feature representations required for recurrent sequence modeling. To overcome this limitation, a hybrid architecture integrating a multi-scale convolutional neural network (CNN), a Transformer, and long short-term memory (LSTM) model operating on short-time Fourier transform (STFT) magnitude representations is introduced. A multi-scale convolutional stem extracts time–frequency features, while frequency-only pooling preserves 29 temporal positions that are represented as feature tokens for Transformer–LSTM processing. Experiments on RadioML 2016.10a across three random seeds and signal-to-noise ratio (SNRs) spanning from −20 to 18 dB show that the proposed model achieves an overall accuracy of 60.46 ± 0.15%, an accuracy of 87.62 ± 0.34% at 0 dB, a mean accuracy of 90.75 ± 0.15% for SNR ≥ 0 dB, and a maximum accuracy of 92.23% at 14 dB. Compared with a multi-scale CNN–LSTM baseline that compresses its features into a single sequence step, the overall, 0 dB, and SNR ≥ 0 dB accuracies are higher by 5.31, 10.26, and 8.80 percentage points, respectively, while the parameter count is reduced by approximately a factor of 9.6 (to 0.90 million). In a single-seed ablation comparison, the 29-step temporal variant achieved overall and 0 dB accuracies that were 3.73 and 6.04 percentage points higher, respectively, than those of the single-step flattened variant. Nevertheless, low-SNR classification and the confusion between WBFM and AM-DSB and between QAM16 and QAM64 remain important challenges.

automatic modulation classification,multi-scale CNN,Transformer,LSTM,RadioML 2016.10a,short-time Fourier transform