This paper introduces a new training strategy to improve speech dereverberation systems using minimal acoustic information and reverberant (wet) speech. Most existing algorithms rely on paired dry/wet data, which is difficult to obtain, or on target metrics that may not adequately capture reverberation characteristics and can lead to poor results on non-target metrics. Our approach uses limited acoustic information, like the reverberation time (RT60), to train a dereverberation system. The system’s output is resynthesized using a generated room impulse response and compared with the original reverberant speech, providing a novel reverberation matching loss replacing the standard target metrics. During inference, only the trained dereverberation model is used. Experimental results demonstrate that our method achieves more consistent performance across various objective metrics used in speech dereverberation than the state-of-the-art.

We introduce O-EENC-SD: an end-to-end online speaker diarization system based on EEND-EDA, featuring a novel RNN-based stitching mechanism for online prediction. In particular, we develop a novel centroid refinement decoder whose usefulness is assessed through a rigorous ablation study. Our system provides key advantages over existing methods: a hyperparameter-free solution compared to unsupervised clustering approaches, and a more efficient alternative to current online end-to-end methods, which are computationally costly. We demonstrate that O-EENC-SD is competitive with the state of the art in the two-speaker conversational telephone speech domain, as tested on the CallHome dataset. Our results show that O-EENC-SD provides a great trade-off between DER and complexity, even when working on independent chunks with no overlap, making the system extremely efficient.

Personalized speech enhancement (PSE) has shown convincing results when it comes to extracting a known target voice among interfering ones. The corresponding systems usually incorporate a representation of the target voice within the enhancement system, which is extracted from an enrollment clip of the target voice with upstream models. Those models are generally heavy as the speaker embedding’s quality directly affects PSE performances. Yet, embeddings generated beforehand cannot account for the variations of the target voice during inference time. In this paper, we propose to perform on-the-fly refinement of the speaker embedding using a tiny speaker encoder. We first introduce a novel contrastive knowledge distillation methodology in order to train a 150k-parameter encoder from complex embeddings. We then use this encoder within the enhancement system during inference and show that the proposed method greatly improves PSE performances while maintaining a low computational load.

Accepted to Interspeech 2025

Generative adversarial network (GAN) models can synthesize high-quality audio signals while ensuring fast sample generation. However, they are difficult to train and are prone to several issues including mode collapse and divergence. In this paper, we introduce SpecDiff-GAN, a neural vocoder based on HiFi-GAN, which was initially devised for speech synthesis from mel spectrogram. In our model, the training stability is enhanced by means of a forward diffusion process which consists in injecting noise from a Gaussian distribution to both real and fake samples before inputting them to the discriminator. We further improve the model by exploiting a spectrally-shaped noise distribution with the aim to make the discriminator's task more challenging. We then show the merits of our proposed model for speech and music synthesis on several datasets. Our experiments confirm that our model compares favorably in audio quality and efficiency compared to several baselines.

Overlapped speech is notoriously problematic for speaker diarization systems. Consequently, the use of speech separation has recently been proposed to improve their performance. Although promising, speech separation models struggle with realistic data because they are trained on simulated mixtures with a fixed number of speakers. In this work, we introduce a new speech separation-guided diarization scheme suitable for the online speaker diarization of long meeting recordings with a variable number of speakers, as present in the AMI corpus. We envisage ConvTasNet and DPRNN as alternatives for the separation networks, with two or three output sources. To obtain the speaker diarization result, voice activity detection is applied on each estimated source. The final model is fine-tuned end-to-end, after first adapting the separation to real data using AMI. The system operates on short segments, and inference is performed by stitching the local predictions using speaker embeddings and incremental clustering. The results show that our system improves the state-of-the-art on the AMI headset mix, using no oracle information and under full evaluation (no collar and including overlapped speech). Finally, we show the strength of our system particularly on overlapped speech sections.

Diffusion models are receiving a growing interest for a variety of signal generation tasks such as speech or music synthesis. WaveGrad, for example, is a successful diffusion model that conditionally uses the mel spectrogram to guide a diffusion process for the generation of high-fidelity audio. However, such models face important challenges concerning the noise diffusion process for training and inference, and they have difficulty generating high-quality speech for speakers that were not seen during training. With the aim of minimizing the conditioning error and increasing the efficiency of the noise diffusion process, we propose in this paper a new scheme called GLA-Grad, which consists in introducing a phase recovery algorithm such as the Griffin-Lim algorithm (GLA) at each step of the regular diffusion process. Furthermore, it can be directly applied to an already-trained waveform generation model, without additional training or fine-tuning. We show that our algorithm outperforms state-of-the-art diffusion models for speech generation, especially when generating speech for a previously unseen target speaker.

Winner-takes-all training is a simple learning paradigm, which handles ambiguous tasks by predicting a set of plausible hypotheses. Recently, a connection was established between Winner-takes-all training and centroidal Voronoi tessellations, showing that, once trained, hypotheses should quantize optimally the shape of the conditional distribution to predict. However, the best use of these hypotheses for uncertainty quantification is still an open question. In this work, we show how to leverage the appealing geometric properties of the Winner-takes-all learners for conditional density estimation, without modifying its original training scheme. We theoretically establish the advantages of our novel estimator both in terms of quantization and density estimation, and we demonstrate its competitiveness on synthetic and real-world datasets, including audio data.

This paper describes a method for estimating the room impulse response (RIR) for a microphone and a sound source located at arbitrary positions from the 3D mesh data of the room.Simulating realistic RIRs with pure physics-driven methods often fails the balance between physical consistency and computational efficiency, hindering application to real-time speech processing.Alternatively, one can use MESH2IR, a fast black-box estimator that consists of an encoder extracting latent code from mesh data with a graph convolutional network (GCN) and a decoder generating the RIR from the latent code.Combining these two approaches, we propose a fast yet physically coherent estimator with interpretable latent code based on differentiable digital signal processing (DDSP).Specifically, the encoder estimates a virtual shoebox room scene that acoustically approximates the real scene, accelerating physical simulation with the differentiable image-source model in the decoder.Our experiments showed that our method outperformed MESH2IR for real mesh data obtained with the depth scanner of Microsoft HoloLens 2, and can provide correct spatial consistency for binaural RIRs.

We introduce Resilient Multiple Choice Learning (rMCL), an extension of the MCL approach for conditional distribution estimation in regression settings where multiple targets may be sampled for each training input. Multiple Choice Learning is a simple framework to tackle multimodal density estimation, using the Winner-Takes-All (WTA) loss for a set of hypotheses. In regression settings, the existing MCL variants focus on merging the hypotheses, thereby eventually sacrificing the diversity of the predictions. In contrast, our method relies on a novel learned scoring scheme underpinned by a mathematical framework based on Voronoi tessellations of the output space, from which we can derive a probabilistic interpretation. After empirically validating rMCL with experiments on synthetic data, we further assess its merits on the sound source localization problem, demonstrating its practical usefulness and the relevance of its interpretation.

This paper describes a blind source separation method for multichannel audio signals, called NF-FastMNMF, based on the integration of the normalizing flow (NF) into the multichannel nonnegative matrix factorization with jointly-diagonalizable spatial covariance matrices, a.k.a. FastMNMF. Whereas the NF of flow-based independent vector analysis, called NF-IVA, acts as the demixing matrices to transform an M-channel mixture into M independent sources, the NF of NF- FastMNMF acts as the diagonalization matrices to transform an M- channel mixture into a spatially-independent M-channel mixture represented as a weighted sum of N source images. This diagonalization enables the NF, which has been used only for determined separation because of its bijective nature, to be applicable to non-determined separation. NF-FastMNMF has time-varying diagonalization matrices that are potentially better at handling dynamical data variation than the time-invariant ones in FastMNMF. To have an NF with richer expression capability, the dimension-wise scalings using diagonal matrices originally used in NF-IVA are replaced with linear transformations using upper triangular matrices; in both cases, the diagonal and upper triangular matrices are estimated by neural networks. The evaluation shows that NF-FastMNMF performs well for both determined and non-determined separations of multiple speech utterances by stationary or non-stationary speakers from a noisy reverberant mixture.

This paper describes noisy speech recognition for an augmented reality headset that helps verbal communication with in real multiparty conversational environments.A major approach that has actively been studied in simulated environments is to sequentially perform speech enhancement and automatic speech recognition (ASR) based on deep neural networks (DNNs) trained in a supervised manner.In our task, however, such a pretrained system fails to work due to the mismatch between the training and test conditions and the head movements of the user.To enhance only the utterances of a target speaker, we use beamforming based on a DNN-based speech mask estimator that can adaptively extract the speech components corresponding to a head-relative particular direction.We propose a semi-supervised adaptation method that jointly updates the mask estimator and the ASR model at run-time using clean speech signals with ground-truth transcriptions and noisy speech signals with highly-confident estimated transcriptions.Comparative experiments using the state-of-theart distant speech recognition system show that the proposed method significantly improves the ASR performance.

This paper describes the practical response- and performance-aware development of online speech enhancement for an augmented reality (AR) headset that helps a user understand conversations made in real noisy echoic environments (e.g., cocktail party). One may use a state-of-the-art blind source separation method called fast multichannel nonnegative matrix factorization (FastMNMF) that works well in various environments thanks to its unsupervised nature. Its heavy computational cost, however, prevents its application to real-time processing. In contrast, a supervised beamforming method that uses a deep neural network (DNN) for estimating spatial information of speech and noise readily fits real-time processing, but suffers from drastic performance degradation in mismatched conditions. Given such complementary characteristics, we propose a dual-process robust online speech enhancement method based on DNN-based beamforming with FastMNMF-guided adaptation. FastMNMF (back end) is performed in a mini-batch style and the noisy and enhanced speech pairs are used together with the original parallel training data for updating the direction-aware DNN (front end) with backpropagation at a computationally-allowable interval. This method is used with a blind dereverberation method called weighted prediction error (WPE) for transcribing the noisy reverberant speech of a speaker, which can be detected from video or selected by a user's hand gesture or eye gaze, in a streaming manner and spatially showing the transcriptions with an AR technique. Our experiment showed that the word error rate was improved by more than 10 points with the run-time adaptation using only twelve minutes observation.

This paper describes heavy-tailed extensions of a state-of-the-art versatile blind source separation method called fast multichannel nonnegative matrix factorization (FastMNMF) from a unified point of view. The common way of deriving such an extension is to replace the multivariate complex Gaussian distribution in the likelihood function with its heavy-tailed generalization,e.g., the multivariate complex Student’s$t$and leptokurtic generalized Gaussian distributions, and tailor-make the corresponding parameter optimization algorithm. Using a wider class of heavy-tailed distributions called a Gaussian scale mixture (GSM),i.e., a mixture of Gaussian distributions whose variances are perturbed by positive random scalars called impulse variables, we propose GSM-FastMNMF and develop an expectation-maximization algorithm that works even when the probability density function of the impulse variables have no analytical expressions. We show that existing heavy-tailed FastMNMF extensions are instances of GSM-FastMNMF and derive a new instance based on the generalized hyperbolic distribution that include the normal-inverse Gaussian, Student’s$t$, and Gaussian distributions as the special cases. Our experiments show that the normal-inverse Gaussian FastMNMF outperforms the state-of-the-art FastMNMF extensions and ILRMA model in speech enhancement and separation in terms of the signal-to-distortion ratio.

This paper describes a computationally-efficient statistical approach to joint (semi-)blind source separation and dereverberation for multichannel noisy reverberant mixture signals. A standard approach to source separation is to formulate a generative model of a multichannel mixture spectrogram that consists of source and spatial models representing the time-frequency power spectral densities (PSDs) and spatial covariance matrices (SCMs) of source images, respectively, and find the maximum-likelihood estimates of these parameters. A state-of-the-art blind source separation method in this thread of research is fast multichannel nonnegative matrix factorization (FastMNMF) based on the low-rank PSDs and jointly-diagonalizable full-rank SCMs. To perform mutually-dependent separation and dereverberation jointly, in this paper we integrate both moving average (MA) and autoregressive (AR) models that represent the early reflections and late reverberations of sources, respectively, into the FastMNMF formalism. Using a pretrained deep generative model of speech PSDs as a source model, we realize semi-blind joint speech separation and dereverberation. We derive an iterative optimization algorithm based on iterative projection or iterative source steering for jointly and efficiently updating the AR parameters and the SCMs. Our experimental results showed the superiority of the proposed ARMA extension over its AR- or MA-ablated version in a speech separation and/or dereverberation task.

This paper describes a joint blind source separation and dereverberation method that works adaptively and efficiently in a reverberant noisy environment. The modern approach to blind source separation (BSS) is to formulate a probabilistic model of multichannel mixture signals that consists of a source model representing the time-frequency structures of source spectrograms and a spatial model representing the inter-channel covariance structures of source images. The cutting-edge BSS method in this thread of research is fast multi-channel nonnegative matrix factorization (FastMNMF) that consists of a low-rank source model based on nonnegative matrix factorization (NMF) and a full-rank spatial model based on jointly-diagonalizable spatial covariance matrices. Although FastMNMF is computationally efficient and can deal with both directional sources and diffuse noise simultaneously, its performance is severely degraded in a reverberant environment. To solve this problem, we propose autoregressive FastMNMF (AR-FastMNMF) based on a unified probabilistic model that combines FastMNMF with a blind dereverberation method called weighted prediction error (WPE), where all the parameters are optimized jointly such that the likelihood for observed reverberant mixture signals is maximized. Experimental results showed the superiority of AR-FastMNMF over conventional methods that perform blind dereverberation and BSS jointly or sequentially.

This paper describes aneural blind source separation (BSS) method based on amortized variational inference (AVI) of a non-linear generative model of mixture signals. A classical statistical approach to BSS is to fit a linear generative model that consists of spatial and source models representing the inter-channel covariances and power spectral densities of sources, respectively. Although the variational autoencoder (VAE) has successfully been used as a non-linear source model with latent features, it should be pretrained from a sufficient amount of isolated signals. Our method, in contrast, enables the VAE-based source model to be trained only from mixture signals. Specifically, we introduce a neural mixture-to-feature inference model that directly infers the latent features from the observed mixture and integrate it with a neural feature-to-mixture generative model consisting of a full-rank spatial model and a VAE-based source model. All the models are optimized jointly such that the likelihood for the training mixtures is maximized in the framework of AVI. Once the inference model is optimized, it can be used for estimating the latent features of sources included in unseen mixture signals. The experimental results show that the proposed method outperformed the state-of-the-art BSS methods based on linear generative models and was comparable to a method based on supervised learning of the VAE-based sourcemodel.

This letter describes a time-varying extension of independent vector analysis (IVA) based on the normalizing flow (NF), called NF-IVA, for determined blind source separation of multichannel audio signals. As in IVA, NF-IVA estimates demixing matrices that transform mixture spectra to source spectra in the complex-valued spatial domain such that the likelihood of those matrices for the mixture spectra is maximized under some non-Gaussian source model. While IVA performs a time-invariant bijective linear transformation, NF-IVA performs a series of time-varying bijective linear transformations (flow blocks) adaptively predicted by neural networks. To regularize such transformations, we introduce a soft volume-preserving (VP) constraint. Given mixture spectra, the parameters of NF-IVA are optimized by gradient descent with backpropagation in an unsupervised manner. Experimental results show that NF-IVA successfully performs speech separation in reverberant environments with different numbers of speakers and microphones and that NF-IVA with the VP constraint outperforms NF-IVA without it, standard IVA with iterative projection, and improved IVA with gradient descent.