Based on TDNN-BiLSTM and Multi-head Attention for Speech Recognition

Effective fusion of local spectral features and global contextual information remains a challenging task in automatic speech recognition (ASR). This paper proposes a hybrid neural network architecture based on time-delay neural networks (TDNN), bidirectional long short-term memory networks (BiLSTM), and multi-head attention mechanisms to improve phoneme sequence prediction accuracy. The model first extracts local acoustic features using TDNN, captures bidirectional temporal dependencies via BiLSTM, and dynamically integrates contextual information through multi-head attention. Experimental results show that the phoneme recognition accuracy of the model is 82.2% and 92.5%, and the word error rate (WER) is 30.9% and 14.1% on the TIMIT and Switchboard test sets, which are 19.3% and 16.5% higher than the traditional TDNN model, and the WER reduction is increased by 23.4% and 33.8%. At the same time, compared with multiple models such as CTC-Transformer, the key role of multi-head attention in modeling long-term dependencies is verified, and the generalization ability of the model for ASR tasks is highlighted.