藉由壓縮性之頻譜損失函數以學習DEMUCS語音強化模型之初步研究

Exploiting the compressed spectral loss for the learning of the DEMUCS speech enhancement network

本研究針對著名的DEMUCS語音強化模型、藉由修改其訓練時所需的損失函數，來提升其效能。DEMUCS由Facebook團隊開發，主要由卷積層組成其編碼模組與解碼模組，而兩模組之間則以長短時記憶模型來對編碼模組之輸出加以分解或降噪。雖然DEMUCS是一個純時域處理的語音強化架構，其訓練所使用的損失函數，卻同時涵蓋了時域和頻域的特徵，其中頻域上的特徵即為訊號經短時間傅立葉轉換所得的頻譜。我們探討當DEMUCS之損失函數中的頻譜其強度值做壓縮時，對於所訓練而得的模型其效能是否有明顯的改變，我們採用的壓縮運算主要是對頻譜強度取一個小於一的正冪次方值，或對頻譜強度取其對數值。當在VoiceBank-DEMAND之資料集上進行評估實驗時，初步結果表明，上述之壓縮運算為取正冪次方值時，其損失函數能使所學習的DEMUCS模型比原DEMUCS模型更有效地提升測試語音的客觀品質與可讀性指標(PESQ與STOI)，充分顯示引入次方壓縮性的頻譜強度於損失函數中能得到語音強化效能更佳的DEMUCS模型。相較而言，當壓縮運算為對數函數時，則沒有改進的效果。

This study aims to improve a highly effective speech enhancement technique, DEMUCS, by revising the respective loss function in learning. DEMUCS, developed by Facebook Team, is built on the Wave-UNet and consists of convolutional layer encoding and decoding blocks with an LSTM layer in between. Although DEMUCS processes the input speech utterance purely in the time (wave) domain, the applied loss function consists of wavedomain L1 distance and multi-scale shorttime- Fourier-transform (STFT) loss. That is, both time- and frequency-domain features are taken into consideration in the learning of DEMUCS. In this study, we present revising the STFT loss in DEMUCS by employing the compressed magnitude spectrogram. The compression is done by either the powerlaw operation with a positive exponent less than one, or the logarithmic operation. We evaluate the presented novel framework on the VoiceBank-DEMAND database and task. The preliminary experimental results suggest that DEMUCS containing the power-law compressed magnitude spectral loss outperforms the original DEMUCS by providing the test utterances with higher objective quality and intelligibility scores (PESQ and STOI). Relatively, the logarithm compressed magnitude spectral loss does not benefit DEMUCS. Therefore, we reveal that DEMUCS can be further improved by properly revising the STFT terms of its loss function.