Research on Voice Pattern Recognition Model for Variable-Duration Speech in Conference Scenario

The traditional method of recording conferences mainly relies on manual labor, resulting in higher time and labor costs. If emerging technology enables sound pattern recognition, it can greatly improve efficiency. Sound pattern recognition can identify speaker identity by extracting the speaker's sound texture characteristics. However, this process encounters specific challenges in practical applications: the variability in speaking duration among different speakers may lead to information loss in acoustic details when extracting acoustic pattern characteristics, thus negatively impacting the model's identification performance. Curren tly, the Transformer model is commonly used for sound pattern recognition. However, due to the residual connection inheritance from the ResNet model, it exhibits a significant decrease in accuracy for long-term voice compared to short-term voice accuracy. Therefore, a VLONGV model combining the Transformer model and convolutional neural network (CNN) blocks is proposed to enhance accuracy in long voice scenarios. The Equal Error Rates (EER) trained by the Transformer model and the VLONGV model 0 n the train-ciean-IOO subset and SITW dataset from the LibriSpeech dataset were 7.13% and 9.2%, and 2.03% and 2.13%, respectively.