Personalized PercepNet: Real-time, Low-complexity Target Voice Separation and Enhancement

Abstract

The presence of multiple talkers in the surrounding environment poses a difficult challenge for real-time speech communication systems considering the constraints on network size and complexity. In this paper, we present Personalized Percep-Net, a real-time speech enhancement model that separates a target speaker from a noisy multi-talker mixture without compromising on complexity of the recently proposed PercepNet. To enable speaker-dependent speech enhancement, we first show how we can train a perceptually motivated speaker embedder network to produce a representative embedding vector for the given speaker. Personalized PercepNet uses the target speaker embedding as additional information to pick out and enhance only the target speaker while suppressing all other competing sounds. Our experiments show that the proposed model significantly outperforms PercepNet and other baselines, both in terms of objective speech enhancement metrics and human opinion scores.

Reference-Based Speech Enhancement via Feature Alignment and Fusion Network

“…Meanwhile, for speech separation, i.e., separating the voice of a target speaker from multi-speaker signals, some works (Wang et al 2018;Mun et al 2020) introduce reference speeches (spoken by the same speaker as that of the target speech) in the form of global embeddings since the global embeddings of target speech and reference speech are correlated. Recently, Giri et al (Giri et al 2021) proposed to utilize the speaker identity embeddings extracted from a clean reference for both speech separation and enhancement. However, there is still no work introducing reference speeches for speech enhancement by exploring local correlations.…”

“…Third, we give results by utilizing the reference in the way of global embedding, similar to (Giri et al 2021;Wang et al 2018). Specifically, we extract the speaker's identity features from the reference speech with a pretrained speaker encoder (Jia et al 2018), and then embed the global vector into our encoder features by concatenating along the feature dimension.…”

“…To further improve the enhancement results, some methods introduce semantic information to assist speech enhancement (Ephrat et al 2018;Hou et al 2018). For example, the speaker identity (ID) is used in speech enhancement (Chuang et al 2019;Giri et al 2021;Shon, Tang, and Glass 2019) and separation tasks (Wang et al 2018;Mun et al 2020;Giri et al 2021), proving that speaker prior derived from the noisy speech or another clean speech spoken by the same ID (i.e., reference speech) is beneficial for the personalized speech enhancement and separation. However, the speaker ID is generally embedded into a global vector, and inserted into the reconstruction network for the enhancement…”

One model to enhance them all: array geometry agnostic multi-channel personalized speech enhancement

“…The speaker embeddings have also been employed for SE. A perceptually motivated PSE model with low complexity was proposed in [2]. [1] introduced two real-time PSE models and tested with reverberant target speech corrupted by both noise and interfering speech.…”

“…Most modern telecommunication services are equipped with a causal/real-time speech enhancement (SE) front-end to deliver high-quality speech audio in noisy environments. Recently, "personalized" SE methods are emerging in the research field by utilizing an enrollment utterance of a target speaker as additional information to not only suppress the ambient noise and reverberation but also remove interfering speech [1,2]. With the ability to handle overlapped speech, personalized speech enhancement (PSE) models significantly improve the perceptual speech quality and the performance of downstream tasks such as automatic speech recognition (ASR) [1].…”