Paris Smaragdis

Zhepei Wang, Cem Subakan, Xilin Jiang et al. · uw

In this paper, we work on a sound recognition system that continually incorporates new sound classes. Our main goal is to develop a framework where the model can be updated without relying on labeled data. For this purpose, we propose adopting representation learning, where an encoder is trained using unlabeled data. This learning framework enables the study and implementation of a practically relevant use case where only a small amount of the labels is available in a continual learning context. We also make the empirical observation that a similarity-based representation learning method within this framework is robust to forgetting even if no explicit mechanism against forgetting is employed. We show that this approach obtains similar performance compared to several distillation-based continual learning methods when employed on self-supervised representation learning methods.

Dimitrios Bralios, Efthymios Tzinis, Gordon Wichern et al.

Traditional source separation approaches train deep neural network models end-to-end with all the data available at once by minimizing the empirical risk on the whole training set. On the inference side, after training the model, the user fetches a static computation graph and runs the full model on some specified observed mixture signal to get the estimated source signals. Additionally, many of those models consist of several basic processing blocks which are applied sequentially. We argue that we can significantly increase resource efficiency during both training and inference stages by reformulating a model's training and inference procedures as iterative mappings of latent signal representations. First, we can apply the same processing block more than once on its output to refine the input signal and consequently improve parameter efficiency. During training, we can follow a block-wise procedure which enables a reduction on memory requirements. Thus, one can train a very complicated network structure using significantly less computation compared to end-to-end training. During inference, we can dynamically adjust how many processing blocks and iterations of a specific block an input signal needs using a gating module.

Francesco Paissan, Luca Della Libera, Zhepei Wang et al.

In this paper, we explore audio-editing with non-rigid text edits. We show that the proposed editing pipeline is able to create audio edits that remain faithful to the input audio. We explore text prompts that perform addition, style transfer, and in-painting. We quantitatively and qualitatively show that the edits are able to obtain results which outperform Audio-LDM, a recently released text-prompted audio generation model. Qualitative inspection of the results points out that the edits given by our approach remain more faithful to the input audio in terms of keeping the original onsets and offsets of the audio events.

Ahmed Mustafa, Jean-Marc Valin, Jan Büthe et al.

GAN vocoders are currently one of the state-of-the-art methods for building high-quality neural waveform generative models. However, most of their architectures require dozens of billion floating-point operations per second (GFLOPS) to generate speech waveforms in samplewise manner. This makes GAN vocoders still challenging to run on normal CPUs without accelerators or parallel computers. In this work, we propose a new architecture for GAN vocoders that mainly depends on recurrent and fully-connected networks to directly generate the time domain signal in framewise manner. This results in considerable reduction of the computational cost and enables very fast generation on both GPUs and low-complexity CPUs. Experimental results show that our Framewise WaveGAN vocoder achieves significantly higher quality than auto-regressive maximum-likelihood vocoders such as LPCNet at a very low complexity of 1.2 GFLOPS. This makes GAN vocoders more practical on edge and low-power devices.

Efthymios Tzinis, Gordon Wichern, Paris Smaragdis et al.

Recent research has shown remarkable performance in leveraging multiple extraneous conditional and non-mutually exclusive semantic concepts for sound source separation, allowing the flexibility to extract a given target source based on multiple different queries. In this work, we propose a new optimal condition training (OCT) method for single-channel target source separation, based on greedy parameter updates using the highest performing condition among equivalent conditions associated with a given target source. Our experiments show that the complementary information carried by the diverse semantic concepts significantly helps to disentangle and isolate sources of interest much more efficiently compared to single-conditioned models. Moreover, we propose a variation of OCT with condition refinement, in which an initial conditional vector is adapted to the given mixture and transformed to a more amenable representation for target source extraction. We showcase the effectiveness of OCT on diverse source separation experiments where it improves upon permutation invariant models with oracle assignment and obtains state-of-the-art performance in the more challenging task of text-based source separation, outperforming even dedicated text-only conditioned models.

Efthymios Tzinis, Gordon Wichern, Aswin Subramanian et al.

We introduce a new paradigm for single-channel target source separation where the sources of interest can be distinguished using non-mutually exclusive concepts (e.g., loudness, gender, language, spatial location, etc). Our proposed heterogeneous separation framework can seamlessly leverage datasets with large distribution shifts and learn cross-domain representations under a variety of concepts used as conditioning. Our experiments show that training separation models with heterogeneous conditions facilitates the generalization to new concepts with unseen out-of-domain data while also performing substantially higher than single-domain specialist models. Notably, such training leads to more robust learning of new harder source separation discriminative concepts and can yield improvements over permutation invariant training with oracle source selection. We analyze the intrinsic behavior of source separation training with heterogeneous metadata and propose ways to alleviate emerging problems with challenging separation conditions. We release the collection of preparation recipes for all datasets used to further promote research towards this challenging task.

Tiago Tavares, Fabio Ayres, Zhepei Wang et al.

Recent advances in audio-text cross-modal contrastive learning have shown its potential towards zero-shot learning. One possibility for this is by projecting item embeddings from pre-trained backbone neural networks into a cross-modal space in which item similarity can be calculated in either domain. This process relies on a strong unimodal pre-training of the backbone networks, and on a data-intensive training task for the projectors. These two processes can be biased by unintentional data leakage, which can arise from using supervised learning in pre-training or from inadvertently training the cross-modal projection using labels from the zero-shot learning evaluation. In this study, we show that a significant part of the measured zero-shot learning accuracy is due to strengths inherited from the audio and text backbones, that is, they are not learned in the cross-modal domain and are not transferred from one modality to another.

Dimitrios Bralios, Efthymios Tzinis, Paris Smaragdis

Recent approaches in source separation leverage semantic information about their input mixtures and constituent sources that when used in conditional separation models can achieve impressive performance. Most approaches along these lines have focused on simple descriptions, which are not always useful for varying types of input mixtures. In this work, we present an approach in which a model, given an input mixture and partial semantic information about a target source, is trained to extract additional semantic data. We then leverage this pre-trained model to improve the separation performance of an uncoupled multi-conditional separation network. Our experiments demonstrate that the separation performance of this multi-conditional model is significantly improved, approaching the performance of an oracle model with complete semantic information. Furthermore, our approach achieves performance levels that are comparable to those of the best performing specialized single conditional models, thus providing an easier to use alternative.

Yutong Wen, Minje Kim, Paris Smaragdis

Music source separation (MSS) aims to extract individual instrument sources from their mixture. While most existing methods focus on the widely adopted four-stem separation setup (vocals, bass, drums, and other instruments), this approach lacks the flexibility needed for real-world applications. To address this, we propose GuideSep, a diffusion-based MSS model capable of instrument-agnostic separation beyond the four-stem setup. GuideSep is conditioned on multiple inputs: a waveform mimicry condition, which can be easily provided by humming or playing the target melody, and mel-spectrogram domain masks, which offer additional guidance for separation. Unlike prior approaches that relied on fixed class labels or sound queries, our conditioning scheme, coupled with the generative approach, provides greater flexibility and applicability. Additionally, we design a mask-prediction baseline using the same model architecture to systematically compare predictive and generative approaches. Our objective and subjective evaluations demonstrate that GuideSep achieves high-quality separation while enabling more versatile instrument extraction, highlighting the potential of user participation in the diffusion-based generative process for MSS. Our code and demo page are available at https://yutongwen.github.io/GuideSep/

Krishna Subramani, Jean-Marc Valin, Umut Isik et al.

Neural vocoders have recently demonstrated high quality speech synthesis, but typically require a high computational complexity. LPCNet was proposed as a way to reduce the complexity of neural synthesis by using linear prediction (LP) to assist an autoregressive model. At inference time, LPCNet relies on the LP coefficients being explicitly computed from the input acoustic features. That makes the design of LPCNet-based systems more complicated, while adding the constraint that the input features must represent a clean speech spectrum. We propose an end-to-end version of LPCNet that lifts these limitations by learning to infer the LP coefficients from the input features in the frame rate network. Results show that the proposed end-to-end approach equals or exceeds the quality of the original LPCNet model, but without explicit LP analysis. Our open-source end-to-end model still benefits from LPCNet's low complexity, while allowing for any type of conditioning features.

Jean-Marc Valin, Umut Isik, Paris Smaragdis et al.

Neural speech synthesis models can synthesize high quality speech but typically require a high computational complexity to do so. In previous work, we introduced LPCNet, which uses linear prediction to significantly reduce the complexity of neural synthesis. In this work, we further improve the efficiency of LPCNet -- targeting both algorithmic and computational improvements -- to make it usable on a wide variety of devices. We demonstrate an improvement in synthesis quality while operating 2.5x faster. The resulting open-source LPCNet algorithm can perform real-time neural synthesis on most existing phones and is even usable in some embedded devices.

Dimitrios Bralios, Paris Smaragdis, Jonah Casebeer

Neural audio autoencoders create compact latent representations that preserve perceptually important information, serving as the foundation for both modern audio compression systems and generation approaches like next-token prediction and latent diffusion. Despite their prevalence, most audio processing operations, such as spatial and spectral up-sampling, still inefficiently operate on raw waveforms or spectral representations rather than directly on these compressed representations. We propose a framework that performs audio processing operations entirely within an autoencoder's latent space, eliminating the need to decode to raw audio formats. Our approach dramatically simplifies training by operating solely in the latent domain, with a latent L1 reconstruction term, augmented by a single latent adversarial discriminator. This contrasts sharply with raw-audio methods that typically require complex combinations of multi-scale losses and discriminators. Through experiments in bandwidth extension and mono-to-stereo up-mixing, we demonstrate computational efficiency gains of up to 100x while maintaining quality comparable to post-processing on raw audio. This work establishes a more efficient paradigm for audio processing pipelines that already incorporate autoencoders, enabling significantly faster and more resource-efficient workflows across various audio tasks.

Dimitrios Bralios, Jonah Casebeer, Paris Smaragdis

Neural audio codecs and autoencoders have emerged as versatile models for audio compression, transmission, feature-extraction, and latent-space generation. However, a key limitation is that most are trained to maximize reconstruction fidelity, often neglecting the specific latent structure necessary for optimal performance in diverse downstream applications. We propose a simple, post-hoc framework to address this by modifying the bottleneck of a pre-trained autoencoder. Our method introduces a "Re-Bottleneck", an inner bottleneck trained exclusively through latent space losses to instill user-defined structure. We demonstrate the framework's effectiveness in three experiments. First, we enforce an ordering on latent channels without sacrificing reconstruction quality. Second, we align latents with semantic embeddings, analyzing the impact on downstream diffusion modeling. Third, we introduce equivariance, ensuring that a filtering operation on the input waveform directly corresponds to a specific transformation in the latent space. Ultimately, our Re-Bottleneck framework offers a flexible and efficient way to tailor representations of neural audio models, enabling them to seamlessly meet the varied demands of different applications with minimal additional training.

Tiago F. Tavares, Fabio Ayres, Paris Smaragdis

Deep Learning techniques have excelled at generating embedding spaces that capture semantic similarities between items. Often these representations are paired, enabling experiments with analogies (pairs within the same domain) and cross-modality (pairs across domains). These experiments are based on specific assumptions about the geometry of embedding spaces, which allow finding paired items by extrapolating the positional relationships between embedding pairs in the training dataset, allowing for tasks such as finding new analogies, and multimodal zero-shot classification. In this work, we propose a metric to evaluate the similarity between paired item representations. Our proposal is built from the structural similarity between the nearest-neighbors induced graphs of each representation, and can be configured to compare spaces based on different distance metrics and on different neighborhood sizes. We demonstrate that our proposal can be used to identify similar structures at different scales, which is hard to achieve with kernel methods such as Centered Kernel Alignment (CKA). We further illustrate our method with two case studies: an analogy task using GloVe embeddings, and zero-shot classification in the CIFAR-100 dataset using CLIP embeddings. Our results show that accuracy in both analogy and zero-shot classification tasks correlates with the embedding similarity. These findings can help explain performance differences in these tasks, and may lead to improved design of paired-embedding models in the future.

Krishna Subramani, Paris Smaragdis, Takuya Higuchi et al.

Non-negative Matrix Factorization (NMF) is a powerful technique for analyzing regularly-sampled data, i.e., data that can be stored in a matrix. For audio, this has led to numerous applications using time-frequency (TF) representations like the Short-Time Fourier Transform. However extending these applications to irregularly-spaced TF representations, like the Constant-Q transform, wavelets, or sinusoidal analysis models, has not been possible since these representations cannot be directly stored in matrix form. In this paper, we formulate NMF in terms of learnable functions (instead of vectors) and show that NMF can be extended to a wider variety of signal classes that need not be regularly sampled.

Karim Helwani, Masahito Togami, Paris Smaragdis et al.

While neural network approaches have made significant strides in resolving classical signal processing problems, it is often the case that hybrid approaches that draw insight from both signal processing and neural networks produce more complete solutions. In this paper, we present a hybrid classical digital signal processing/deep neural network (DSP/DNN) approach to source separation (SS) highlighting the theoretical link between variational autoencoder and classical approaches to SS. We propose a system that transforms the single channel under-determined SS task to an equivalent multichannel over-determined SS problem in a properly designed latent space. The separation task in the latent space is treated as finding a variational block-wise disentangled representation of the mixture. We show empirically, that the design choices and the variational formulation of the task at hand motivated by the classical signal processing theoretical results lead to robustness to unseen out-of-distribution data and reduction of the overfitting risk. To address the resulting permutation issue we explicitly incorporate a novel differentiable permutation loss function and augment the model with a memory mechanism to keep track of the statistics of the individual sources.

Cameron Churchwell, Minje Kim, Paris Smaragdis

Selecting appropriate inductive biases is an essential step in the design of machine learning models, especially when working with audio, where even short clips may contain millions of samples. To this end, we propose the combolutional layer: a learned-delay IIR comb filter and fused envelope detector, which extracts harmonic features in the time domain. We demonstrate the efficacy of the combolutional layer on three information retrieval tasks, evaluate its computational cost relative to other audio frontends, and provide efficient implementations for training. We find that the combolutional layer is an effective replacement for convolutional layers in audio tasks where precise harmonic analysis is important, e.g., piano transcription, speaker classification, and key detection. Additionally, the combolutional layer has several other key benefits over existing frontends, namely: low parameter count, efficient CPU inference, strictly real-valued computations, and improved interpretability.

Shayan K. Azmoodeh, Krishna Subramani, Paris Smaragdis

We generalize the low-rank decomposition problem, such as principal and independent component analysis (PCA, ICA) for continuous-time vector-valued signals and provide a model-agnostic implicit neural signal representation framework to learn numerical approximations to solve the problem. Modeling signals as continuous-time stochastic processes, we unify the approaches to both the PCA and ICA problems in the continuous setting through a contrast function term in the network loss, enforcing the desired statistical properties of the source signals (decorrelation, independence) learned in the decomposition. This extension to a continuous domain allows the application of such decompositions to point clouds and irregularly sampled signals where standard techniques are not applicable.

Riccardo Miccini, Minje Kim, Clément Laroche et al.

Speech enhancement (SE) enables robust speech recognition, real-time communication, hearing aids, and other applications where speech quality is crucial. However, deploying such systems on resource-constrained devices involves choosing a static trade-off between performance and computational efficiency. In this paper, we introduce dynamic slimming to DEMUCS, a popular SE architecture, making it scalable and input-adaptive. Slimming lets the model operate at different utilization factors (UF), each corresponding to a different performance/efficiency trade-off, effectively mimicking multiple model sizes without the extra storage costs. In addition, a router subnet, trained end-to-end with the backbone, determines the optimal UF for the current input. Thus, the system saves resources by adaptively selecting smaller UFs when additional complexity is unnecessary. We show that our solution is Pareto-optimal against individual UFs, confirming the benefits of dynamic routing. When training the proposed dynamically-slimmable model to use 10% of its capacity on average, we obtain the same or better speech quality as the equivalent static 25% utilization while reducing MACs by 29%.

Zhepei Wang, Cem Subakan, Krishna Subramani et al.

Recent advances in using language models to obtain cross-modal audio-text representations have overcome the limitations of conventional training approaches that use predefined labels. This has allowed the community to make progress in tasks like zero-shot classification, which would otherwise not be possible. However, learning such representations requires a large amount of human-annotated audio-text pairs. In this paper, we study unsupervised approaches to improve the learning framework of such representations with unpaired text and audio. We explore domain-unspecific and domain-specific curation methods to create audio-text pairs that we use to further improve the model. We also show that when domain-specific curation is used in conjunction with a soft-labeled contrastive loss, we are able to obtain significant improvement in terms of zero-shot classification performance on downstream sound event classification or acoustic scene classification tasks.

Efthymios Tzinis, Yossi Adi, Vamsi Krishna Ithapu et al.

We present RemixIT, a simple yet effective self-supervised method for training speech enhancement without the need of a single isolated in-domain speech nor a noise waveform. Our approach overcomes limitations of previous methods which make them dependent on clean in-domain target signals and thus, sensitive to any domain mismatch between train and test samples. RemixIT is based on a continuous self-training scheme in which a pre-trained teacher model on out-of-domain data infers estimated pseudo-target signals for in-domain mixtures. Then, by permuting the estimated clean and noise signals and remixing them together, we generate a new set of bootstrapped mixtures and corresponding pseudo-targets which are used to train the student network. Vice-versa, the teacher periodically refines its estimates using the updated parameters of the latest student models. Experimental results on multiple speech enhancement datasets and tasks not only show the superiority of our method over prior approaches but also showcase that RemixIT can be combined with any separation model as well as be applied towards any semi-supervised and unsupervised domain adaptation task. Our analysis, paired with empirical evidence, sheds light on the inside functioning of our self-training scheme wherein the student model keeps obtaining better performance while observing severely degraded pseudo-targets.

Jonah Casebeer, Nicholas J. Bryan, Paris Smaragdis

Adaptive filtering algorithms are commonplace in signal processing and have wide-ranging applications from single-channel denoising to multi-channel acoustic echo cancellation and adaptive beamforming. Such algorithms typically operate via specialized online, iterative optimization methods and have achieved tremendous success, but require expert knowledge, are slow to develop, and are difficult to customize. In our work, we present a new method to automatically learn adaptive filtering update rules directly from data. To do so, we frame adaptive filtering as a differentiable operator and train a learned optimizer to output a gradient descent-based update rule from data via backpropagation through time. We demonstrate our general approach on an acoustic echo cancellation task (single-talk with noise) and show that we can learn high-performing adaptive filters for a variety of common linear and non-linear multidelayed block frequency domain filter architectures. We also find that our learned update rules exhibit fast convergence, can optimize in the presence of nonlinearities, and are robust to acoustic scene changes despite never encountering any during training.

Zhepei Wang, Jonah Casebeer, Adam Clemmitt et al.

In this work, we present HIDACT, a novel network architecture for adaptive computation for efficiently recognizing acoustic events. We evaluate the model on a sound event detection task where we train it to adaptively process frequency bands. The model learns to adapt to the input without requesting all frequency sub-bands provided. It can make confident predictions within fewer processing steps, hence reducing the amount of computation. Experimental results show that HIDACT has comparable performance to baseline models with more parameters and higher computational complexity. Furthermore, the model can adjust the amount of computation based on the data and computational budget.

Marco A. Martínez Ramírez, Oliver Wang, Paris Smaragdis et al.

We present a data-driven approach to automate audio signal processing by incorporating stateful third-party, audio effects as layers within a deep neural network. We then train a deep encoder to analyze input audio and control effect parameters to perform the desired signal manipulation, requiring only input-target paired audio data as supervision. To train our network with non-differentiable black-box effects layers, we use a fast, parallel stochastic gradient approximation scheme within a standard auto differentiation graph, yielding efficient end-to-end backpropagation. We demonstrate the power of our approach with three separate automatic audio production applications: tube amplifier emulation, automatic removal of breaths and pops from voice recordings, and automatic music mastering. We validate our results with a subjective listening test, showing our approach not only can enable new automatic audio effects tasks, but can yield results comparable to a specialized, state-of-the-art commercial solution for music mastering.

Efthymios Tzinis, Jonah Casebeer, Zhepei Wang et al.

We propose FEDENHANCE, an unsupervised federated learning (FL) approach for speech enhancement and separation with non-IID distributed data across multiple clients. We simulate a real-world scenario where each client only has access to a few noisy recordings from a limited and disjoint number of speakers (hence non-IID). Each client trains their model in isolation using mixture invariant training while periodically providing updates to a central server. Our experiments show that our approach achieves competitive enhancement performance compared to IID training on a single device and that we can further facilitate the convergence speed and the overall performance using transfer learning on the server-side. Moreover, we show that we can effectively combine updates from clients trained locally with supervised and unsupervised losses. We also release a new dataset LibriFSD50K and its creation recipe in order to facilitate FL research for source separation problems.

Krishna Subramani, Paris Smaragdis

Most audio processing pipelines involve transformations that act on fixed-dimensional input representations of audio. For example, when using the Short Time Fourier Transform (STFT) the DFT size specifies a fixed dimension for the input representation. As a consequence, most audio machine learning models are designed to process fixed-size vector inputs which often prohibits the repurposing of learned models on audio with different sampling rates or alternative representations. We note, however, that the intrinsic spectral information in the audio signal is invariant to the choice of the input representation or the sampling rate. Motivated by this, we introduce a novel way of processing audio signals by treating them as a collection of points in feature space, and we use point cloud machine learning models that give us invariance to the choice of representation parameters, such as DFT size or the sampling rate. Additionally, we observe that these methods result in smaller models, and allow us to significantly subsample the input representation with minimal effects to a trained model performance.

Efthymios Tzinis, Zhepei Wang, Xilin Jiang et al.

Recent progress in audio source separation lead by deep learning has enabled many neural network models to provide robust solutions to this fundamental estimation problem. In this study, we provide a family of efficient neural network architectures for general purpose audio source separation while focusing on multiple computational aspects that hinder the application of neural networks in real-world scenarios. The backbone structure of this convolutional network is the SUccessive DOwnsampling and Resampling of Multi-Resolution Features (SuDoRM-RF) as well as their aggregation which is performed through simple one-dimensional convolutions. This mechanism enables our models to obtain high fidelity signal separation in a wide variety of settings where variable number of sources are present and with limited computational resources (e.g. floating point operations, memory footprint, number of parameters and latency). Our experiments show that SuDoRM-RF models perform comparably and even surpass several state-of-the-art benchmarks with significantly higher computational resource requirements. The causal variation of SuDoRM-RF is able to obtain competitive performance in real-time speech separation of around 10dB scale-invariant signal-to-distortion ratio improvement (SI-SDRi) while remaining up to 20 times faster than real-time on a laptop device.

Jonah Casebeer, Jamshed Kaikaus, Paris Smaragdis

In this paper, we present a method for jointly-learning a microphone selection mechanism and a speech enhancement network for multi-channel speech enhancement with an ad-hoc microphone array. The attention-based microphone selection mechanism is trained to reduce communication-costs through a penalty term which represents a task-performance/ communication-cost trade-off. While working within the trade-off, our method can intelligently stream from more microphones in lower SNR scenes and fewer microphones in higher SNR scenes. We evaluate the model in complex echoic acoustic scenes with moving sources and show that it matches the performance of models that stream from a fixed number of microphones while reducing communication costs.

An Zhao, Krishna Subramani, Paris Smaragdis

The Short-Time Fourier Transform (STFT) has been a staple of signal processing, often being the first step for many audio tasks. A very familiar process when using the STFT is the search for the best STFT parameters, as they often have significant side effects if chosen poorly. These parameters are often defined in terms of an integer number of samples, which makes their optimization non-trivial. In this paper we show an approach that allows us to obtain a gradient for STFT parameters with respect to arbitrary cost functions, and thus enable the ability to employ gradient descent optimization of quantities like the STFT window length, or the STFT hop size. We do so for parameter values that stay constant throughout an input, but also for cases where these parameters have to dynamically change over time to accommodate varying signal characteristics.

Efthymios Tzinis, Dimitrios Bralios, Paris Smaragdis

Recent deep learning approaches have shown great improvement in audio source separation tasks. However, the vast majority of such work is focused on improving average separation performance, often neglecting to examine or control the distribution of the results. In this paper, we propose a simple, unified gradient reweighting scheme, with a lightweight modification to bias the learning process of a model and steer it towards a certain distribution of results. More specifically, we reweight the gradient updates of each batch, using a user-specified probability distribution. We apply this method to various source separation tasks, in order to shift the operating point of the models towards different objectives. We demonstrate different parameterizations of our unified reweighting scheme can be used towards addressing several real-world problems, such as unreliable separation estimates. Our framework enables the user to control a robustness trade-off between worst and average performance. Moreover, we experimentally show that our unified reweighting scheme can also be used in order to shift the focus of the model towards being more accurate for user-specified sound classes or even towards easier examples in order to enable faster convergence.

Efthymios Tzinis, Zhepei Wang, Paris Smaragdis

In this paper, we present an efficient neural network for end-to-end general purpose audio source separation. Specifically, the backbone structure of this convolutional network is the SUccessive DOwnsampling and Resampling of Multi-Resolution Features (SuDoRMRF) as well as their aggregation which is performed through simple one-dimensional convolutions. In this way, we are able to obtain high quality audio source separation with limited number of floating point operations, memory requirements, number of parameters and latency. Our experiments on both speech and environmental sound separation datasets show that SuDoRMRF performs comparably and even surpasses various state-of-the-art approaches with significantly higher computational resource requirements.

Yu-Che Wang, Shrikant Venkataramani, Paris Smaragdis

Supervised learning for single-channel speech enhancement requires carefully labeled training examples where the noisy mixture is input into the network and the network is trained to produce an output close to the ideal target. To relax the conditions on the training data, we consider the task of training speech enhancement networks in a self-supervised manner. We first use a limited training set of clean speech sounds and learn a latent representation by autoencoding on their magnitude spectrograms. We then autoencode on speech mixtures recorded in noisy environments and train the resulting autoencoder to share a latent representation with the clean examples. We show that using this training schema, we can now map noisy speech to its clean version using a network that is autonomously trainable without requiring labeled training examples or human intervention.

Shrikant Venkataramani, Efthymios Tzinis, Paris Smaragdis

Discriminative models for source separation have recently been shown to produce impressive results. However, when operating on sources outside of the training set, these models can not perform as well and are cumbersome to update. Classical methods like Non-negative Matrix Factorization (NMF) provide modular approaches to source separation that can be easily updated to adapt to new mixture scenarios. In this paper, we generalize NMF to develop end-to-end non-negative auto-encoders and demonstrate how they can be used for source separation. Our experiments indicate that these models deliver comparable separation performance to discriminative approaches, while retaining the modularity of NMF and the modeling flexibility of neural networks.

Efthymios Tzinis, Shrikant Venkataramani, Zhepei Wang et al.

In this paper, we propose a two-step training procedure for source separation via a deep neural network. In the first step we learn a transform (and it's inverse) to a latent space where masking-based separation performance using oracles is optimal. For the second step, we train a separation module that operates on the previously learned space. In order to do so, we also make use of a scale-invariant signal to distortion ratio (SI-SDR) loss function that works in the latent space, and we prove that it lower-bounds the SI-SDR in the time domain. We run various sound separation experiments that show how this approach can obtain better performance as compared to systems that learn the transform and the separation module jointly. The proposed methodology is general enough to be applicable to a large class of neural network end-to-end separation systems.

Zhepei Wang, Cem Subakan, Efthymios Tzinis et al.

Continual learning consists in incrementally training a model on a sequence of datasets and testing on the union of all datasets. In this paper, we examine continual learning for the problem of sound classification, in which we wish to refine already trained models to learn new sound classes. In practice one does not want to maintain all past training data and retrain from scratch, but naively updating a model with new data(sets) results in a degradation of already learned tasks, which is referred to as "catastrophic forgetting." We develop a generative replay procedure for generating training audio spectrogram data, in place of keeping older training datasets. We show that by incrementally refining a classifier with generative replay a generator that is 4% of the size of all previous training data matches the performance of refining the classifier keeping 20% of all previous training data. We thus conclude that we can extend a trained sound classifier to learn new classes without having to keep previously used datasets.

Jonah Casebeer, Michael Colomb, Paris Smaragdis

We propose a completely unsupervised method to understand audio scenes observed with random microphone arrangements by decomposing the scene into its constituent sources and their relative presence in each microphone. To this end, we formulate a neural network architecture that can be interpreted as a nonnegative tensor factorization of a multi-channel audio recording. By clustering on the learned network parameters corresponding to channel content, we can learn sources' individual spectral dictionaries and their activation patterns over time. Our method allows us to leverage deep learning advances like end-to-end training, while also allowing stochastic minibatch training so that we can feasibly decompose realistic audio scenes that are intractable to decompose using standard methods. This neural network architecture is easily extensible to other kinds of tensor factorizations.

Shrikant Venkataramani, Efthymios Tzinis, Paris Smaragdis

Training neural networks for source separation involves presenting a mixture recording at the input of the network and updating network parameters in order to produce an output that resembles the clean source. Consequently, supervised source separation depends on the availability of paired mixture-clean training examples. In this paper, we interpret source separation as a style transfer problem. We present a variational auto-encoder network that exploits the commonality across the domain of mixtures and the domain of clean sounds and learns a shared latent representation across the two domains. Using these cycle-consistent variational auto-encoders, we learn a mapping from the mixture domain to the domain of clean sounds and perform source separation without explicitly supervising with paired training examples.

Efthymios Tzinis, Shrikant Venkataramani, Paris Smaragdis

We present a monophonic source separation system that is trained by only observing mixtures with no ground truth separation information. We use a deep clustering approach which trains on multi-channel mixtures and learns to project spectrogram bins to source clusters that correlate with various spatial features. We show that using such a training process we can obtain separation performance that is as good as making use of ground truth separation information. Once trained, this system is capable of performing sound separation on monophonic inputs, despite having learned how to do so using multi-channel recordings.

Jonah Casebeer, Zhepei Wang, Paris Smaragdis

In this paper we introduce the idea of multi-view networks for sound classification with multiple sensors. We show how one can build a multi-channel sound recognition model trained on a fixed number of channels, and deploy it to scenarios with arbitrary (and potentially dynamically changing) number of input channels and not observe degradation in performance. We demonstrate that at inference time you can safely provide this model all available channels as it can ignore noisy information and leverage new information better than standard baseline approaches. The model is evaluated in both an anechoic environment and in rooms generated by a room acoustics simulator. We demonstrate that this model can generalize to unseen numbers of channels as well as unseen room geometries.

Shrikant Venkataramani, Paris Smaragdis

The performance of single channel source separation algorithms has improved greatly in recent times with the development and deployment of neural networks. However, many such networks continue to operate on the magnitude spectrogram of a mixture, and produce an estimate of source magnitude spectrograms, to perform source separation. In this paper, we interpret these steps as additional neural network layers and propose an end-to-end source separation network that allows us to estimate the separated speech waveform by operating directly on the raw waveform of the mixture. Furthermore, we also propose the use of masking based end-to-end separation networks that jointly optimize the mask and the latent representations of the mixture waveforms. These networks show a significant improvement in separation performance compared to existing architectures in our experiments. To train these end-to-end models, we investigate the use of composite cost functions that are derived from objective evaluation metrics as measured on waveforms. We present subjective listening test results that demonstrate the improvement attained by using masking based end-to-end networks and also reveal insights into the performance of these cost functions for end-to-end source separation.

Jonah Casebeer, Brian Luc, Paris Smaragdis

We propose a set of denoising neural networks capable of operating on an arbitrary number of channels at runtime, irrespective of how many channels they were trained on. We coin the proposed models multi-view networks since they operate using multiple views of the same data. We explore two such architectures and show how they outperform traditional denoising models in multi-channel scenarios. Additionally, we demonstrate how multi-view networks can leverage information provided by additional recordings to make better predictions, and how they are able to generalize to a number of recordings not seen in training.

Shrikant Venkataramani, Ryley Higa, Paris Smaragdis

Recent neural network strategies for source separation attempt to model audio signals by processing their waveforms directly. Mean squared error (MSE) that measures the Euclidean distance between waveforms of denoised speech and the ground-truth speech, has been a natural cost-function for these approaches. However, MSE is not a perceptually motivated measure and may result in large perceptual discrepancies. In this paper, we propose and experiment with new loss functions for end-to-end source separation. These loss functions are motivated by BSS\_Eval and perceptual metrics like source to distortion ratio (SDR), source to interference ratio (SIR), source to artifact ratio (SAR) and short-time objective intelligibility ratio (STOI). This enables the flexibility to mix and match these loss functions depending upon the requirements of the task. Subjective listening tests reveal that combinations of the proposed cost functions help achieve superior separation performance as compared to stand-alone MSE and SDR costs.

Cem Subakan, Oluwasanmi Koyejo, Paris Smaragdis

Popular generative model learning methods such as Generative Adversarial Networks (GANs), and Variational Autoencoders (VAE) enforce the latent representation to follow simple distributions such as isotropic Gaussian. In this paper, we argue that learning a complicated distribution over the latent space of an auto-encoder enables more accurate modeling of complicated data distributions. Based on this observation, we propose a two stage optimization procedure which maximizes an approximate implicit density model. We experimentally verify that our method outperforms GANs and VAEs on two image datasets (MNIST, CELEB-A). We also show that our approach is amenable to learning generative model for sequential data, by learning to generate speech and music.

Cem Subakan, Paris Smaragdis

Generative source separation methods such as non-negative matrix factorization (NMF) or auto-encoders, rely on the assumption of an output probability density. Generative Adversarial Networks (GANs) can learn data distributions without needing a parametric assumption on the output density. We show on a speech source separation experiment that, a multi-layer perceptron trained with a Wasserstein-GAN formulation outperforms NMF, auto-encoders trained with maximum likelihood, and variational auto-encoders in terms of source to distortion ratio.

Shrikant Venkataramani, Y. Cem Subakan, Paris Smaragdis

Convolutive Non-Negative Matrix Factorization model factorizes a given audio spectrogram using frequency templates with a temporal dimension. In this paper, we present a convolutional auto-encoder model that acts as a neural network alternative to convolutive NMF. Using the modeling flexibility granted by neural networks, we also explore the idea of using a Recurrent Neural Network in the encoder. Experimental results on speech mixtures from TIMIT dataset indicate that the convolutive architecture provides a significant improvement in separation performance in terms of BSSeval metrics.

Nasser Mohammadiha, Paris Smaragdis, Arne Leijon

Reducing the interference noise in a monaural noisy speech signal has been a challenging task for many years. Compared to traditional unsupervised speech enhancement methods, e.g., Wiener filtering, supervised approaches, such as algorithms based on hidden Markov models (HMM), lead to higher-quality enhanced speech signals. However, the main practical difficulty of these approaches is that for each noise type a model is required to be trained a priori. In this paper, we investigate a new class of supervised speech denoising algorithms using nonnegative matrix factorization (NMF). We propose a novel speech enhancement method that is based on a Bayesian formulation of NMF (BNMF). To circumvent the mismatch problem between the training and testing stages, we propose two solutions. First, we use an HMM in combination with BNMF (BNMF-HMM) to derive a minimum mean square error (MMSE) estimator for the speech signal with no information about the underlying noise type. Second, we suggest a scheme to learn the required noise BNMF model online, which is then used to develop an unsupervised speech enhancement system. Extensive experiments are carried out to investigate the performance of the proposed methods under different conditions. Moreover, we compare the performance of the developed algorithms with state-of-the-art speech enhancement schemes using various objective measures. Our simulations show that the proposed BNMF-based methods outperform the competing algorithms substantially.

Nasser Mohammadiha, Paris Smaragdis, Ghazaleh Panahandeh et al.

Nonnegative matrix factorization (NMF) has been actively investigated and used in a wide range of problems in the past decade. A significant amount of attention has been given to develop NMF algorithms that are suitable to model time series with strong temporal dependencies. In this paper, we propose a novel state-space approach to perform dynamic NMF (D-NMF). In the proposed probabilistic framework, the NMF coefficients act as the state variables and their dynamics are modeled using a multi-lag nonnegative vector autoregressive (N-VAR) model within the process equation. We use expectation maximization and propose a maximum-likelihood estimation framework to estimate the basis matrix and the N-VAR model parameters. Interestingly, the N-VAR model parameters are obtained by simply applying NMF. Moreover, we derive a maximum a posteriori estimate of the state variables (i.e., the NMF coefficients) that is based on a prediction step and an update step, similarly to the Kalman filter. We illustrate the benefits of the proposed approach using different numerical simulations where D-NMF significantly outperforms its static counterpart. Experimental results for three different applications show that the proposed approach outperforms two state-of-the-art NMF approaches that exploit temporal dependencies, namely a nonnegative hidden Markov model and a frame stacking approach, while it requires less memory and computational power.

Shrikant Venkataramani, Jonah Casebeer, Paris Smaragdis

Source separation and other audio applications have traditionally relied on the use of short-time Fourier transforms as a front-end frequency domain representation step. The unavailability of a neural network equivalent to forward and inverse transforms hinders the implementation of end-to-end learning systems for these applications. We present an auto-encoder neural network that can act as an equivalent to short-time front-end transforms. We demonstrate the ability of the network to learn optimal, real-valued basis functions directly from the raw waveform of a signal and further show how it can be used as an adaptive front-end for supervised source separation. In terms of separation performance, these transforms significantly outperform their Fourier counterparts. Finally, we also propose a novel source to distortion ratio based cost function for end-to-end source separation.

Y. Cem Subakan, Paris Smaragdis

In this paper, we propose a new Recurrent Neural Network (RNN) architecture. The novelty is simple: We use diagonal recurrent matrices instead of full. This results in better test likelihood and faster convergence compared to regular full RNNs in most of our experiments. We show the benefits of using diagonal recurrent matrices with popularly used LSTM and GRU architectures as well as with the vanilla RNN architecture, on four standard symbolic music datasets.

Mohammad Babaeizadeh, Paris Smaragdis, Roy H. Campbell

Neural networks are usually over-parameterized with significant redundancy in the number of required neurons which results in unnecessary computation and memory usage at inference time. One common approach to address this issue is to prune these big networks by removing extra neurons and parameters while maintaining the accuracy. In this paper, we propose NoiseOut, a fully automated pruning algorithm based on the correlation between activations of neurons in the hidden layers. We prove that adding additional output neurons with entirely random targets results into a higher correlation between neurons which makes pruning by NoiseOut even more efficient. Finally, we test our method on various networks and datasets. These experiments exhibit high pruning rates while maintaining the accuracy of the original network.