Few-Shot Speaker Identification Using Lightweight Prototypical Network With Feature Grouping and Interaction

Existing methods for few-shot speaker identification (FSSI) obtain high accuracy, but their computational complexities and model sizes need to be reduced lightweight applications. In this work, we propose a FSSI method using prototypical network with the final goal implement on intelligent terminals limited resources, such as smart watches speakers. proposed network, an embedding module is designed perform feature grouping reducing memory requirement complexity, interaction enhancing representational ability of learned embedding. module, audio each speech sample split into several low-dimensional subsets that are transformed by recurrent convolutional block in parallel. Then, operations averaging, addition, concatenation, element-wise summation statistics pooling sequentially executed learn sample. The consists bidirectional long short-term memory, de-redundancy convolution which conducted too. Our compared baseline three datasets selected from public corpora (VoxCeleb1, VoxCeleb2, LibriSpeech). results show our obtains higher accuracy under conditions, has advantages over all complexity size.