A study on phonemes recognition method for Mandarin pronunciation based on improved Zipformer-RNN-T(Pruned) modeling

Abstract

In recent years, empowered by artificial intelligence technologies, computer-assisted language learning systems have gradually become a hot topic of research. Currently, the mainstream pronunciation assessment models rely on advanced speech recognition technology, converting speech into phoneme sequences, and then determining mispronounced phonemes through sequence comparison. To optimize the phoneme recognition task in pronunciation evaluation, this paper proposes a Chinese pronunciation phoneme recognition model based on the improved Zipformer-RNN-T(Pruned) architecture, aiming to improve recognition accuracy and reduce parameter count. First, the AISHELL1-PHONEME and ST-CMDS-PHONEME datasets for Mandarin phoneme recognition through data preprocessing. Then, three layers of the Zipformer Block architecture are introduced into the Zipformer encoder to significantly enhance model performance. In the stateless Pred Network, the GELU activation function is adopted to effectively prevent neuron deactivation. Furthermore, a hybrid Pruned RNN-T/CTC Loss fusion strategy is proposed, further optimizing recognition performance. The experimental results demonstrate that the method performs excellently in the phoneme recognition task, achieving a Word Error Rate (WER) of 1.92% (Dev) and 2.12% (Test) on the AISHELL1-PHONEME dataset, and 4.28% (Dev) and 4.51% (Test) on the ST-CMDS-PHONEME dataset. Moreover, the model requires only 61.1M parameters, striking a balance between performance and efficiency.