Conventional speech signal keyword extraction often uses graph neural network algorithms，which achieves keyword extraction through the representation of key information feature vectors. However，this method lacks recognition of speech signal scale feature components，resulting in poor quality of final keyword extraction. A speech signal keyword extraction method based on autoregressive pre trained language model is proposed. Based on the waveform changes of the speech signal in the time domain，the signal is denoised，and the scale feature components are obtained by multi-modal decomposition of the signal. From this，the word vector features of the speech signal are constructed，and the semantic vector parameters of the entire speech signal are obtained by combining clustering algorithms. The semantic vector parameters of the initially selected keywords are then deduplicated. An autoregressive pre trained language model is introduced to calculate the similarity between candidate keywords and semantic vectors of speech signals. The keyword extraction of speech signals are achieved. The experimental results show that within the range of 5~30 keywords，the extraction recall of the proposed method remains above 80%. The proposed method can effectively improve the quality of extracting keywords from speech signals，is easy to implement，and can be widely applied in the field of speech signal processing.