Xiaofei Li

Xian Li, Nian Shao, Xiaofei Li

Self-supervised learning (SSL) has emerged as a popular approach for learning audio representations. One goal of audio self-supervised pre-training is to transfer knowledge to downstream audio tasks, generally including clip-level and frame-level tasks. While frame-level tasks are important for fine-grained acoustic scene/event understanding, prior studies primarily evaluate on clip-level downstream tasks. In order to tackle both clip-level and frame-level tasks, this paper proposes Audio Teacher-Student Transformer (ATST), with a clip-level version (named ATST-Clip) and a frame-level version (named ATST-Frame), responsible for learning clip-level and frame-level representations, respectively. Both methods use a Transformer encoder and a teacher-student training scheme. We have carefully designed the view creation strategy for ATST-Clip and ATST-Frame. Specifically, ATST-Clip uses segment-wise data augmentations, and ATST-Frame integrates frame-wise data augmentations and masking. Experimental results show that our ATST-Frame model obtains state-of-the-art (SOTA) performances on most of the clip-level and frame-level downstream tasks. Especially, it outperforms other models by a large margin on the frame-level sound event detection task. In addition, the performance can be further improved by combining the two models through knowledge distillation. Our code is available online.

Xian Li, Xiaofei Li

Self-supervised learning (SSL) learns knowledge from a large amount of unlabeled data, and then transfers the knowledge to a specific problem with a limited number of labeled data. SSL has achieved promising results in various domains. This work addresses the problem of segment-level general audio SSL, and proposes a new transformer-based teacher-student SSL model, named ATST. A transformer encoder is developed on a recently emerged teacher-student baseline scheme, which largely improves the modeling capability of pre-training. In addition, a new strategy for positive pair creation is designed to fully leverage the capability of transformer. Extensive experiments have been conducted, and the proposed model achieves the new state-of-the-art results on almost all of the downstream tasks.

Yujie Yang, Changsheng Quan, Xiaofei Li

In multichannel speech enhancement, both spectral and spatial information are vital for discriminating between speech and noise. How to fully exploit these two types of information and their temporal dynamics remains an interesting research problem. As a solution to this problem, this paper proposes a multi-cue fusion network named McNet, which cascades four modules to respectively exploit the full-band spatial, narrow-band spatial, sub-band spectral, and full-band spectral information. Experiments show that each module in the proposed network has its unique contribution and, as a whole, notably outperforms other state-of-the-art methods.

Nian Shao, Rui Zhou, Pengyu Wang et al.

In this work, we propose CleanMel, a single-channel Mel-spectrogram denoising and dereverberation network for improving both speech quality and automatic speech recognition (ASR) performance. The proposed network takes as input the noisy and reverberant microphone recording and predicts the corresponding clean Mel-spectrogram. The enhanced Mel-spectrogram can be either transformed to the speech waveform with a neural vocoder or directly used for ASR. The proposed network is composed of interleaved cross-band and narrow-band processing in the Mel-frequency domain, for learning the full-band spectral pattern and the narrow-band properties of signals, respectively. Compared to linear-frequency domain or time-domain speech enhancement, the key advantage of Mel-spectrogram enhancement is that Mel-frequency presents speech in a more compact way and thus is easier to learn, which will benefit both speech quality and ASR. Experimental results on five English and one Chinese datasets demonstrate a significant improvement in both speech quality and ASR performance achieved by the proposed model.Code and audio examples of our model are available online.

Bing Yang, Hong Liu, Xiaofei Li

Multiple moving sound source localization in real-world scenarios remains a challenging issue due to interaction between sources, time-varying trajectories, distorted spatial cues, etc. In this work, we propose to use deep learning techniques to learn competing and time-varying direct-path phase differences for localizing multiple moving sound sources. A causal convolutional recurrent neural network is designed to extract the direct-path phase difference sequence from signals of each microphone pair. To avoid the assignment ambiguity and the problem of uncertain output-dimension encountered when simultaneously predicting multiple targets, the learning target is designed in a weighted sum format, which encodes source activity in the weight and direct-path phase differences in the summed value. The learned direct-path phase differences for all microphone pairs can be directly used to construct the spatial spectrum according to the formulation of steered response power (SRP). This deep neural network (DNN) based SRP method is referred to as SRP-DNN. The locations of sources are estimated by iteratively detecting and removing the dominant source from the spatial spectrum, in which way the interaction between sources is reduced. Experimental results on both simulated and real-world data show the superiority of the proposed method in the presence of noise and reverberation.

Bing Yang, Hong Liu, Xiaofei Li

Direct-path relative transfer function (DP-RTF) refers to the ratio between the direct-path acoustic transfer functions of two microphone channels. Though DP-RTF fully encodes the sound spatial cues and serves as a reliable localization feature, it is often erroneously estimated in the presence of noise and reverberation. This paper proposes to learn DP-RTF with deep neural networks for robust binaural sound source localization. A DP-RTF learning network is designed to regress the binaural sensor signals to a real-valued representation of DP-RTF. It consists of a branched convolutional neural network module to separately extract the inter-channel magnitude and phase patterns, and a convolutional recurrent neural network module for joint feature learning. To better explore the speech spectra to aid the DP-RTF estimation, a monaural speech enhancement network is used to recover the direct-path spectrograms from the noisy ones. The enhanced spectrograms are stacked onto the noisy spectrograms to act as the input of the DP-RTF learning network. We train one unique DP-RTF learning network using many different binaural arrays to enable the generalization of DP-RTF learning across arrays. This way avoids time-consuming training data collection and network retraining for a new array, which is very useful in practical application. Experimental results on both simulated and real-world data show the effectiveness of the proposed method for direction of arrival (DOA) estimation in the noisy and reverberant environment, and a good generalization ability to unseen binaural arrays.

Nian Shao, Erfan Loweimi, Xiaofei Li

Sound event detection (SED), as a core module of acoustic environmental analysis, suffers from the problem of data deficiency. The integration of semi-supervised learning (SSL) largely mitigates such problem while bringing no extra annotation budget. This paper researches on several core modules of SSL, and introduces a random consistency training (RCT) strategy. First, a self-consistency loss is proposed to fuse with the teacher-student model to stabilize the training. Second, a hard mixup data augmentation is proposed to account for the additive property of sounds. Third, a random augmentation scheme is applied to flexibly combine different types of data augmentations. Experiments show that the proposed strategy outperform other widely-used strategies.

Changsheng Quan, Xiaofei Li

This paper addresses the problem of multi-channel multi-speech separation based on deep learning techniques. In the short time Fourier transform domain, we propose an end-to-end narrow-band network that directly takes as input the multi-channel mixture signals of one frequency, and outputs the separated signals of this frequency. In narrow-band, the spatial information (or inter-channel difference) can well discriminate between speakers at different positions. This information is intensively used in many narrow-band speech separation methods, such as beamforming and clustering of spatial vectors. The proposed network is trained to learn a rule to automatically exploit this information and perform speech separation. Such a rule should be valid for any frequency, thence the network is shared by all frequencies. In addition, a full-band permutation invariant training criterion is proposed to solve the frequency permutation problem encountered by most narrow-band methods. Experiments show that, by focusing on deeply learning the narrow-band information, the proposed method outperforms the oracle beamforming method and the state-of-the-art deep learning based method.

Tianwei Zhang, Huayan Zhang, Xiaofei Li et al.

Dynamic objects in the environment, such as people and other agents, lead to challenges for existing simultaneous localization and mapping (SLAM) approaches. To deal with dynamic environments, computer vision researchers usually apply some learning-based object detectors to remove these dynamic objects. However, these object detectors are computationally too expensive for mobile robot on-board processing. In practical applications, these objects output noisy sounds that can be effectively detected by on-board sound source localization. The directional information of the sound source object can be efficiently obtained by direction of sound arrival (DoA) estimation, but depth estimation is difficult. Therefore, in this paper, we propose a novel audio-visual fusion approach that fuses sound source direction into the RGB-D image and thus removes the effect of dynamic obstacles on the multi-robot SLAM system. Experimental results of multi-robot SLAM in different dynamic environments show that the proposed method uses very small computational resources to obtain very stable self-localization results.

Siyuan Zhang, Xiaofei Li

This paper addresses the problem of microphone array generalization for deep-learning-based end-to-end multichannel speech enhancement. We aim to train a unique deep neural network (DNN) potentially performing well on unseen microphone arrays. The microphone array geometry shapes the network's parameters when training on a fixed microphone array, and thus restricts the generalization of the trained network to another microphone array. To resolve this problem, a single network is trained using data recorded by various microphone arrays of different geometries. We design three variants of our recently proposed narrowband network to cope with the agnostic number of microphones. Overall, the goal is to make the network learn the universal information for speech enhancement that is available for any array geometry, rather than learn the one-array-dedicated characteristics. The experiments on both simulated and real room impulse responses (RIR) demonstrate the excellent across-array generalization capability of the proposed networks, in the sense that their performance measures are very close to, or even exceed the network trained with test arrays. Moreover, they notably outperform various beamforming methods and other advanced deep-learning-based methods.

Xiaofei Li

Sound event detection is a core module for acoustic environmental analysis. Semi-supervised learning technique allows to largely scale up the dataset without increasing the annotation budget, and recently attracts lots of research attention. In this work, we study on two advanced semi-supervised learning techniques for sound event detection. Data augmentation is important for the success of recent deep learning systems. This work studies the audio-signal random augmentation method, which provides an augmentation strategy that can handle a large number of different audio transformations. In addition, consistency regularization is widely adopted in recent state-of-the-art semi-supervised learning methods, which exploits the unlabelled data by constraining the prediction of different transformations of one sample to be identical to the prediction of this sample. This work finds that, for semi-supervised sound event detection, consistency regularization is an effective strategy, especially the best performance is achieved when it is combined with the MeanTeacher model.

Xiaofei Li, Radu Horaud

This paper proposes a delayed subband LSTM network for online monaural (single-channel) speech enhancement. The proposed method is developed in the short time Fourier transform (STFT) domain. Online processing requires frame-by-frame signal reception and processing. A paramount feature of the proposed method is that the same LSTM is used across frequencies, which drastically reduces the number of network parameters, the amount of training data and the computational burden. Training is performed in a subband manner: the input consists of one frequency, together with a few context frequencies. The network learns a speech-to-noise discriminative function relying on the signal stationarity and on the local spectral pattern, based on which it predicts a clean-speech mask at each frequency. To exploit future information, i.e. look-ahead, we propose an output-delayed subband architecture, which allows the unidirectional forward network to process a few future frames in addition to the current frame. We leverage the proposed method to participate to the DNS real-time speech enhancement challenge. Experiments with the DNS dataset show that the proposed method achieves better performance-measuring scores than the DNS baseline method, which learns the full-band spectra using a gated recurrent unit network.

Xiaofei LI, Radu Horaud

In this paper, we address the problem of multichannel speech enhancement in the short-time Fourier transform (STFT) domain. A long short-time memory (LSTM) network takes as input a sequence of STFT coefficients associated with a frequency bin of multichannel noisy-speech signals. The network's output is the corresponding sequence of single-channel cleaned speech. We propose several clean-speech network targets, namely, the magnitude ratio mask, the complex STFT coefficients and the (smoothed) spatial filter. A prominent feature of the proposed model is that the same LSTM architecture, with identical parameters, is trained across frequency bins. The proposed method is referred to as narrow-band deep filtering. This choice stays in contrast with traditional wide-band speech enhancement methods. The proposed deep filtering is able to discriminate between speech and noise by exploiting their different temporal and spatial characteristics: speech is non-stationary and spatially coherent while noise is relatively stationary and weakly correlated across channels. This is similar in spirit with unsupervised techniques, such as spectral subtraction and beamforming. We describe extensive experiments with both mixed signals (noise is added to clean speech) and real signals (live recordings). We empirically evaluate the proposed architecture variants using speech enhancement and speech recognition metrics, and we compare our results with the results obtained with several state of the art methods. In the light of these experiments we conclude that narrow-band deep filtering has very good speech enhancement and speech recognition performance, and excellent generalization capabilities in terms of speaker variability and noise type.

Xiaofei Li, Simon Leglaive, Laurent Girin et al.

We propose a method using a long short-term memory (LSTM) network to estimate the noise power spectral density (PSD) of single-channel audio signals represented in the short time Fourier transform (STFT) domain. An LSTM network common to all frequency bands is trained, which processes each frequency band individually by mapping the noisy STFT magnitude sequence to its corresponding noise PSD sequence. Unlike deep-learning-based speech enhancement methods that learn the full-band spectral structure of speech segments, the proposed method exploits the sub-band STFT magnitude evolution of noise with a long time dependency, in the spirit of the unsupervised noise estimators described in the literature. Speaker- and speech-independent experiments with different types of noise show that the proposed method outperforms the unsupervised estimators, and generalizes well to noise types that are not present in the training set.

Xiaofei Li, Yutong Ban, Laurent Girin et al.

We address the problem of online localization and tracking of multiple moving speakers in reverberant environments. The paper has the following contributions. We use the direct-path relative transfer function (DP-RTF), an inter-channel feature that encodes acoustic information robust against reverberation, and we propose an online algorithm well suited for estimating DP-RTFs associated with moving audio sources. Another crucial ingredient of the proposed method is its ability to properly assign DP-RTFs to audio-source directions. Towards this goal, we adopt a maximum-likelihood formulation and we propose to use an exponentiated gradient (EG) to efficiently update source-direction estimates starting from their currently available values. The problem of multiple speaker tracking is computationally intractable because the number of possible associations between observed source directions and physical speakers grows exponentially with time. We adopt a Bayesian framework and we propose a variational approximation of the posterior filtering distribution associated with multiple speaker tracking, as well as an efficient variational expectation-maximization (VEM) solver. The proposed online localization and tracking method is thoroughly evaluated using two datasets that contain recordings performed in real environments.

Xiaofei Li, Radu Horaud, Sharon Gannot

This paper addresses the problems of blind channel identification and multichannel equalization for speech dereverberation and noise reduction. The time-domain cross-relation method is not suitable for blind room impulse response identification, due to the near-common zeros of the long impulse responses. We extend the cross-relation method to the short-time Fourier transform (STFT) domain, in which the time-domain impulse responses are approximately represented by the convolutive transfer functions (CTFs) with much less coefficients. The CTFs suffer from the common zeros caused by the oversampled STFT. We propose to identify CTFs based on the STFT with the oversampled signals and the critical sampled CTFs, which is a good compromise between the frequency aliasing of the signals and the common zeros problem of CTFs. In addition, a normalization of the CTFs is proposed to remove the gain ambiguity across sub-bands. In the STFT domain, the identified CTFs is used for multichannel equalization, in which the sparsity of speech signals is exploited. We propose to perform inverse filtering by minimizing the $\ell_1$-norm of the source signal with the relaxed $\ell_2$-norm fitting error between the micophone signals and the convolution of the estimated source signal and the CTFs used as a constraint. This method is advantageous in that the noise can be reduced by relaxing the $\ell_2$-norm to a tolerance corresponding to the noise power, and the tolerance can be automatically set. The experiments confirm the efficiency of the proposed method even under conditions with high reverberation levels and intense noise.

Israel D. Gebru, Silèye Ba, Xiaofei Li et al.

Speaker diarization consists of assigning speech signals to people engaged in a dialogue. An audio-visual spatiotemporal diarization model is proposed. The model is well suited for challenging scenarios that consist of several participants engaged in multi-party interaction while they move around and turn their heads towards the other participants rather than facing the cameras and the microphones. Multiple-person visual tracking is combined with multiple speech-source localization in order to tackle the speech-to-person association problem. The latter is solved within a novel audio-visual fusion method on the following grounds: binaural spectral features are first extracted from a microphone pair, then a supervised audio-visual alignment technique maps these features onto an image, and finally a semi-supervised clustering method assigns binaural spectral features to visible persons. The main advantage of this method over previous work is that it processes in a principled way speech signals uttered simultaneously by multiple persons. The diarization itself is cast into a latent-variable temporal graphical model that infers speaker identities and speech turns, based on the output of an audio-visual association process, executed at each time slice, and on the dynamics of the diarization variable itself. The proposed formulation yields an efficient exact inference procedure. A novel dataset, that contains audio-visual training data as well as a number of scenarios involving several participants engaged in formal and informal dialogue, is introduced. The proposed method is thoroughly tested and benchmarked with respect to several state-of-the art diarization algorithms.

Xiaofei Li, Laurent Girin, Radu Horaud et al.

This paper addresses the problem of binaural localization of a single speech source in noisy and reverberant environments. For a given binaural microphone setup, the binaural response corresponding to the direct-path propagation of a single source is a function of the source direction. In practice, this response is contaminated by noise and reverberations. The direct-path relative transfer function (DP-RTF) is defined as the ratio between the direct-path acoustic transfer function of the two channels. We propose a method to estimate the DP-RTF from the noisy and reverberant microphone signals in the short-time Fourier transform domain. First, the convolutive transfer function approximation is adopted to accurately represent the impulse response of the sensors in the STFT domain. Second, the DP-RTF is estimated by using the auto- and cross-power spectral densities at each frequency and over multiple frames. In the presence of stationary noise, an inter-frame spectral subtraction algorithm is proposed, which enables to achieve the estimation of noise-free auto- and cross-power spectral densities. Finally, the estimated DP-RTFs are concatenated across frequencies and used as a feature vector for the localization of speech source. Experiments with both simulated and real data show that the proposed localization method performs well, even under severe adverse acoustic conditions, and outperforms state-of-the-art localization methods under most of the acoustic conditions.