Speech Enhancement Method Based on Modified Encoder-Decoder Pyramid Transformer

The development of new technologies for voice communication has led to the need improvement speech enhancement methods. Modern users information systems place high demands on both intelligibility signal and its perceptual quality. In this work we propose a approach solving problem enhancement. For this, modified pyramidal transformer neural network with an encoder-decoder structure was developed. encoder compressed spectrum into series internal embeddings. decoder self-attention transformations reconstructed mask complex ratio cleaned noisy signals based embeddings calculated by encoder. Two possible loss functions were considered training proposed model. It shown that use frequency encoding mixed input data improved performance approach. trained tested DNS Challenge 2021 dataset. showed compared modern We provide qualitative analysis process implemented network. gradually moved from simple noise masking in early epochs restoring missing formant components speaker's later epochs. This metrics subjective quality enhanced speech.