A Single Channel Speech Enhancement Framework Based on Input-Information Enhanced Conformer

Recently, the Convolution-augmented transformer (Conformer) has shown competitive results in speech applications. However, Conformer faces problems such as insufficient signal information, difficulty in processing long sequence information, and poor performance in tasks that require modeling complex temporal dynamics in single channel speech enhancement research. LSTM can have better performance in longer sequences. Stacking several audio frames into one block (SAB) in preprocessing allows the network to read more information once, which can compensate for the insufficient contextual information in single channel speech enhancement. Therefore, this paper proposes a single channel speech enhancement framework based on input-information enhanced Conformer, Nesting Conformer with LSTM, the input signal is first processed by SAB and LSTM before being input into the Conformer module. The experiment shows that compared with previous methods using the same dataset and parameters, the proposed framework's perception evaluation of speech quality (PESQ) increases by 19.0% under different signal-to-noise ratios.