High Fidelity Speech Enhancement with Band-split RNN

This report presents the development of our speech enhancement system, which includes the use of a recently proposed music separation model, the band-split recurrent neural network (BSRNN), and a MetricGAN-based training objective to improve non-differentiable quality metrics such as perceptual evaluation of speech quality (PESQ) score. Experiment conducted on Interspeech 2021 DNS challenge shows that our BSRNN system outperforms various top-ranking benchmark systems in previous deep noise suppression (DNS) challenges and achieves state-of-the-art (SOTA) result on the DNS-2020 non-blind test set in both offline and online scenarios.

Parameter-efficient transfer learning of pre-trained Transformer models for speaker verification using adapters

Recently, the pre-trained Transformer models have received a rising interest in the field of speech processing thanks to their great success in various downstream tasks. However, most fine-tuning approaches update all the parameters of the pre-trained model, which becomes prohibitive as the model size grows and sometimes results in overfitting on small datasets. In this paper, we conduct a comprehensive analysis of applying parameter-efficient transfer learning (PETL) methods to reduce the required learnable parameters for adapting to speaker verification tasks. Specifically, during the fine-tuning process, the pre-trained models are frozen, and only lightweight modules inserted in each Transformer block are trainable (a method known as adapters). Moreover, to boost the performance in a cross-language low-resource scenario, the Transformer model is further tuned on a large intermediate dataset before directly fine-tuning it on a small dataset. With updating fewer than 4% of parameters, (our proposed) PETL-based methods achieve comparable performances with full fine-tuning methods (Vox1-O: 0.55%, Vox1-E: 0.82%, Vox1-H:1.73%).

Improving Dual-Microphone Speech Enhancement by Learning Cross-Channel Features with Multi-Head Attention

Hand-crafted spatial features, such as inter-channel intensity difference (IID) and inter-channel phase difference (IPD), play a fundamental role in recent deep learning based dual-microphone speech enhancement (DMSE) systems. However, learning the mutual relationship between artificially designed spatial and spectral features is hard in the end-to-end DMSE. In this work, a novel architecture for DMSE using a multi-head cross-attention based convolutional recurrent network (MHCA-CRN) is presented. The proposed MHCA-CRN model includes a channel-wise encoding structure for preserving intra-channel features and a multi-head cross-attention mechanism for fully exploiting cross-channel features. In addition, the proposed approach specifically formulates the decoder with an extra SNR estimator to estimate frame-level SNR under a multi-task learning framework, which is expected to avoid speech distortion led by end-to-end DMSE module. Finally, a spectral gain function is adopted to further suppress the unnatural residual noise. Experiment results demonstrated superior performance of the proposed model against several state-of-the-art models.

Speaker-Aware Mixture of Mixtures Training for Weakly Supervised Speaker Extraction

Dominant researches adopt supervised training for speaker extraction, while the scarcity of ideally clean corpus and channel mismatch problem are rarely considered. To this end, we propose speaker-aware mixture of mixtures training (SAMoM), utilizing the consistency of speaker identity among target source, enrollment utterance and target estimate to weakly supervise the training of a deep speaker extractor. In SAMoM, the input is constructed by mixing up different speaker-aware mixtures (SAMs), each contains multiple speakers with their identities known and enrollment utterances available. Informed by enrollment utterances, target speech is extracted from the input one by one, such that the estimated targets can approximate the original SAMs after a remix in accordance with the identity consistency. Moreover, using SAMoM in a semi-supervised setting with a certain amount of clean sources enables application in noisy scenarios. Extensive experiments on Libri2Mix show that the proposed method achieves promising results without access to any clean sources (11.06dB SI-SDRi). With a domain adaptation, our approach even outperformed supervised framework in a cross-domain evaluation on AISHELL-1.

Target Confusion in End-to-end Speaker Extraction: Analysis and Approaches

Recently, end-to-end speaker extraction has attracted increasing attention and shown promising results. However, its performance is often inferior to that of a blind source separation (BSS) counterpart with a similar network architecture, due to the auxiliary speaker encoder may sometimes generate ambiguous speaker embeddings. Such ambiguous guidance information may confuse the separation network and hence lead to wrong extraction results, which deteriorates the overall performance. We refer to this as the target confusion problem. In this paper, we conduct an analysis of such an issue and solve it in two stages. In the training phase, we propose to integrate metric learning methods to improve the distinguishability of embeddings produced by the speaker encoder. While for inference, a novel post-filtering strategy is designed to revise the wrong results. Specifically, we first identify these confusion samples by measuring the similarities between output estimates and enrollment utterances, after which the true target sources are recovered by a subtraction operation. Experiments show that performance improvement of more than 1dB SI-SDRi can be brought, which validates the effectiveness of our methods and emphasizes the impact of the target confusion problem.

Learning Decoupling Features Through Orthogonality Regularization

Keyword spotting (KWS) and speaker verification (SV) are two important tasks in speech applications. Research shows that the state-of-art KWS and SV models are trained independently using different datasets since they expect to learn distinctive acoustic features. However, humans can distinguish language content and the speaker identity simultaneously. Motivated by this, we believe it is important to explore a method that can effectively extract common features while decoupling task-specific features. Bearing this in mind, a two-branch deep network (KWS branch and SV branch) with the same network structure is developed and a novel decoupling feature learning method is proposed to push up the performance of KWS and SV simultaneously where speaker-invariant keyword representations and keyword-invariant speaker representations are expected respectively. Experiments are conducted on Google Speech Commands Dataset (GSCD). The results demonstrate that the orthogonality regularization helps the network to achieve SOTA EER of 1.31% and 1.87% on KWS and SV, respectively.

Text Anchor Based Metric Learning for Small-footprint Keyword Spotting

Keyword Spotting (KWS) remains challenging to achieve the trade-off between small footprint and high accuracy. Recently proposed metric learning approaches improved the generalizability of models for the KWS task, and 1D-CNN based KWS models have achieved the state-of-the-arts (SOTA) in terms of model size. However, for metric learning, due to data limitations, the speech anchor is highly susceptible to the acoustic environment and speakers. Also, we note that the 1D-CNN models have limited capability to capture long-term temporal acoustic features. To address the above problems, we propose to utilize text anchors to improve the stability of anchors. Furthermore, a new type of model (LG-Net) is exquisitely designed to promote long-short term acoustic feature modeling based on 1D-CNN and self-attention. Experiments are conducted on Google Speech Commands Dataset version 1 (GSCDv1) and 2 (GSCDv2). The results demonstrate that the proposed text anchor based metric learning method shows consistent improvements over speech anchor on representative CNN-based models. Moreover, our LG-Net model achieves SOTA accuracy of 97.67% and 96.79% on two datasets, respectively. It is encouraged to see that our lighter LG-Net with only 74k parameters obtains 96.82% KWS accuracy on the GSCDv1 and 95.77% KWS accuracy on the GSCDv2.

Complex Neural Spatial Filter: Enhancing Multi-channel Target Speech Separation in Complex Domain

To date, mainstream target speech separation (TSS) approaches are formulated to estimate the complex ratio mask (cRM) of the target speech in time-frequency domain under supervised deep learning framework. However, the existing deep models for estimating cRM are designed in the way that the real and imaginary parts of the cRM are separately modeled using real-valued training data pairs. The research motivation of this study is to design a deep model that fully exploits the temporal-spectral-spatial information of multi-channel signals for estimating cRM directly and efficiently in complex domain. As a result, a novel TSS network is designed consisting of two modules, a complex neural spatial filter (cNSF) and an MVDR. Essentially, cNSF is a cRM estimation model and an MVDR module is cascaded to the cNSF module to reduce the nonlinear speech distortions introduced by neural network. Specifically, to fit the cRM target, all input features of cNSF are reformulated into complex-valued representations following the supervised learning paradigm. Then, to achieve good hierarchical feature abstraction, a complex deep neural network (cDNN) is delicately designed with U-Net structure. Experiments conducted on simulated multi-channel speech data demonstrate the proposed cNSF outperforms the baseline NSF by 12.1% scale-invariant signal-to-distortion ratio and 33.1% word error rate.

Layer Reduction: Accelerating Conformer-Based Self-Supervised Model via Layer Consistency

Transformer-based self-supervised models are trained as feature extractors and have empowered many downstream speech tasks to achieve state-of-the-art performance. However, both the training and inference process of these models may encounter prohibitively high computational cost and large parameter budget. Although Parameter Sharing Strategy (PSS) proposed in ALBERT paves the way for parameter reduction, the computation required remains the same. Interestingly, we found in experiments that distributions of feature embeddings from different Transformer layers are similar when PSS is integrated: a property termed as Layer Consistency (LC) in this paper. Given this similarity of feature distributions, we assume that feature embeddings from different layers would have similar representing power. In this work, Layer Consistency enables us to adopt Transformer-based models in a more efficient manner: the number of Conformer layers in each training iteration could be uniformly sampled and Shallow Layer Inference (SLI) could be applied to reduce the number of layers in inference stage. In experiments, our models are trained with LibriSpeech dataset and then evaluated on both phone classification and Speech Recognition tasks. We experimentally achieve 7.8X parameter reduction, 41.9% training speedup and 37.7% inference speedup while maintaining comparable performance with conventional BERT-like self-supervised methods.

Temporal-Spatial Neural Filter: Direction Informed End-to-End Multi-channel Target Speech Separation

Target speech separation refers to extracting the target speaker's speech from mixed signals. Despite the recent advances in deep learning based close-talk speech separation, the applications to real-world are still an open issue. Two main challenges are the complex acoustic environment and the real-time processing requirement. To address these challenges, we propose a temporal-spatial neural filter, which directly estimates the target speech waveform from multi-speaker mixture in reverberant environments, assisted with directional information of the speaker(s). Firstly, against variations brought by complex environment, the key idea is to increase the acoustic representation completeness through the jointly modeling of temporal, spectral and spatial discriminability between the target and interference source. Specifically, temporal, spectral, spatial along with the designed directional features are integrated to create a joint acoustic representation. Secondly, to reduce the latency, we design a fully-convolutional autoencoder framework, which is purely end-to-end and single-pass. All the feature computation is implemented by the network layers and operations to speed up the separation procedure. Evaluation is conducted on simulated reverberant dataset WSJ0-2mix and WSJ0-3mix under speaker-independent scenario. Experimental results demonstrate that the proposed method outperforms state-of-the-art deep learning based multi-channel approaches with fewer parameters and faster processing speed. Furthermore, the proposed temporal-spatial neural filter can handle mixtures with varying and unknown number of speakers and exhibits persistent performance even when existing a direction estimation error. Codes and models will be released soon.