基于时间尺度聚合的短语音说话人识别

Speaker recognition distinguishes identities based on individual voice features and is widely used in voice assistants and security systems. Short speech data, due to its limited duration, makes it difficult to extract stable speaker features, which reduces recognition accuracy. Traditional multi-scale feature aggregation methods often focus on information fusion along the channel dimension but may miss important temporal dynamics in short speech scenarios. This paper proposes a multi-scale feature aggregation method based on temporal features. A multi-scale feature extraction module captures both local and global temporal features in short speech. This approach enhances the complementarity of features at different scales, maintaining recognition accuracy even when the model size is reduced by 50%, with an accuracy improvement of approximately 1%.