Dual-Branch Attention-In-Attention Transformer for Single-Channel Speech Enhancement

CMGAN: Conformer-Based Metric-GAN for Monaural Speech Enhancement

“…Since the phase is unstructured and always challenging to estimate, this might result in an inaccurate magnitude estimation to compensate for the challenging phase. This problem can be mitigated by including both complex and magnitude losses or by complex refinement approaches, which basically decouple the problem into estimating a bounded mask for the magnitude followed by a complex refinement branch to further improve the magnitude and estimate the phase from the denoised complex representations [13], [24], [58]- [60]. However, since recent studies recommended mapping-based methods over the preceding masking-based approaches for complex spectrogram estimation [22], [61], the complex refinement branch would follow a mapping-based approach.…”

“…For the time-domain methods, we included the standard SEGAN [14] and three recent methods: TSTNN [16], DEMUCS [17] and SE-Conformer [18]. For the TF-domain methods, we evaluate six recent SOTA methods, i.e., MetricGAN [11], PHASEN [12], PFPL [104], Metric-GAN+ [105], DB-AIAT [13] and DPT-FSNet [26]. It can be observed that most of the TF-domain methods outperform the time-domain counterparts over all utilized metrics.…”

“…SE techniques can be roughly categorized into two prominent families of approaches. Chronologically, enhancing the speech time-frequency (TF) representation (spectrogram) constitutes the classical SE paradigm which encompasses the majority of model-based as well as more recent DL approaches [5], [11]- [13]. More recently, a new set of approaches were introduced to enhance raw speech time-domain waveform directly without any transformational overheads [14]- [18].…”

“…II-A. Recent studies propose enhancing the magnitude followed by complex spectrogram refinement, which can alleviate the compensation problem effectively [13], [24]. Furthermore, the commonly used objective function in SE is simply the L p −norm distance between the estimated and the target spectrograms.…”

“…In addition, many approaches utilize transformers [25] to capture the long-term dependency in the waveform or the spectrogram [13], [16], [26]. Recently, conformers have been introduced as an alternative to transformers in ASR and speech separation tasks due to their capacity of capturing both local context and global context [27], [28].…”