Language-queried audio source separation (LASS) is an emerging research field that has recently received increasing attention. This task aims to isolate individual sources from a mixture of signals using natural language descriptions, enabling applications in various areas such as automatic audio editing. While conventional methods focus on the system architecture, the important aspect of data processing has been overlooked. The data for training LASS are typically created by mixing various audio signals in the dataset to form a mixture. One signal is then used as the target, whereas the others are regarded as interference. However, sound events in the target signal could overlap with those in the interference signals, which may cause confusion that instructs the model to both retain and suppress the same sound events within a single training example. In addition, training LASS with large-scale datasets may suffer from the data imbalance problem, where some sound events appear too frequently while others are rare. In this paper, we address these problems by using data sampling techniques. Specifically, the interference signals are sampled so that their audio tags do not conflict with those of the target signal, where the tags are generated using an audio tagging model. To balance the data, we consider several balanced sampling approaches using tag or caption embedding. By leveraging their distribution information, we use either weighted or group sampling to boost the occurrence of underrepresented samples while reducing the presence of overrepresented ones. Experimental results show the superiority of the proposed method over state-of-the-art LASS systems in DCASE 2024 Challenge Task 9. Pre-trained model is available at: https://github.com/tucothien/LASS-CLBS.

With the rapid development of neural networks in recent years, the ability of various networks to enhance the magnitude spectrum of noisy speech in the single-channel speech enhancement domain has become exceptionally outstanding. However, enhancing the phase spectrum using neural networks is often ineffective, which remains a challenging problem. In this paper, we found that the human ear cannot sensitively perceive the difference between a precise phase spectrum and a biased phase (BP) spectrum. Therefore, we propose an optimization method of phase reconstruction, allowing freedom on the global-phase bias instead of reconstructing the precise phase spectrum. We applied it to a Conformer-based Metric Generative Adversarial Networks (CMGAN) baseline model, which relaxes the existing constraints of precise phase and gives the neural network a broader learning space. Results show that this method achieves a new state-of-the-art performance without incurring additional computational overhead.

Self-supervised learning (SSL) using masked prediction has made great strides in general-purpose audio representation. This study proposes Masked Modeling Duo (M2D), an improved masked prediction SSL, which learns by predicting representations of masked input signals that serve as training signals. Unlike conventional methods, M2D obtains a training signal by encoding only the masked part, encouraging the two networks in M2D to model the input. While M2D improves general-purpose audio representations, a specialized representation is essential for real-world applications, such as in industrial and medical domains. The often confidential and proprietary data in such domains is typically limited in size and has a different distribution from that in pre-training datasets. Therefore, we propose M2D for X (M2D-X), which extends M2D to enable the pre-training of specialized representations for an application X. M2D-X learns from M2D and an additional task and inputs background noise. We make the additional task configurable to serve diverse applications, while the background noise helps learn on small data and forms a denoising task that makes representation robust. With these design choices, M2D-X should learn a representation specialized to serve various application needs. Our experiments confirmed that the representations for general-purpose audio, specialized for the highly competitive AudioSet and speech domain, and a small-data medical task achieve top-level performance, demonstrating the potential of using our models as a universal audio pre-training framework. Our code is available online for future studies.

Masked Autoencoders is a simple yet powerful self-supervised learning method. However, it learns representations indirectly by reconstructing masked input patches. Several methods learn representations directly by predicting representations of masked patches; however, we think using all patches to encode training signal representations is suboptimal. We propose a new method, Masked Modeling Duo (M2D), that learns representations directly while obtaining training signals using only masked patches. In the M2D, the online network encodes visible patches and predicts masked patch representations, and the target network, a momentum encoder, encodes masked patches. To better predict target representations, the online network should model the input well, while the target network should also model it well to agree with online predictions. Then the learned representations should better model the input. We validated the M2D by learning general-purpose audio representations, and M2D set new state-of-the-art performance on tasks such as UrbanSound8K, VoxCeleb1, AudioSet20K, GTZAN, and SpeechCommandsV2.

Pre-trained models are essential as feature extractors in modern machine learning systems in various domains. In this study, we hypothesize that representations effective for general audio tasks should provide multiple aspects of robust features of the input sound. For recognizing sounds regardless of perturbations such as varying pitch or timbre, features should be robust to these perturbations. For serving the diverse needs of tasks such as recognition of emotions or music genres, representations should provide multiple aspects of information, such as local and global features. To implement our principle, we propose a self-supervised learning method: Bootstrap Your Own Latent (BYOL) for Audio (BYOL-A, pronounced “viola”). BYOL-A pre-trains representations of the input sound invariant to audio data augmentations, which makes the learned representations robust to the perturbations of sounds. Whereas the BYOL-A encoder combines local and global features and calculates their statistics to make the representation provide multi-aspect information. As a result, the learned representations should provide robust and multi-aspect information to serve various needs of diverse tasks. We evaluated the general audio task performance of BYOL-A compared to previous state-of-the-art methods, and BYOL-A demonstrated generalizability with the best average result of 72.4% and the best VoxCeleb1 result of 57.6%. Extensive ablation experiments revealed that the BYOL-A encoder architecture contributes to most performance, and the final critical portion resorts to the BYOL framework and BYOL-A augmentations. Our code is available online for future studies.

We propose a novel framework for target speech extraction based on semantic information, called ConceptBeam. Target speech extraction means extracting the speech of a target speaker in a mixture. Typical approaches have been exploiting properties of audio signals, such as harmonic structure and direction of arrival. In contrast, ConceptBeam tackles the problem with semantic clues. Specifically, we extract the speech of speakers speaking about a concept, i.e., a topic of interest, using a concept specifier such as an image or speech. Solving this novel problem would open the door to innovative applications such as listening systems that focus on a particular topic discussed in a conversation. Unlike keywords, concepts are abstract notions, making it challenging to directly represent a target concept. In our scheme, a concept is encoded as a semantic embedding by mapping the concept specifier to a shared embedding space. This modality-independent space can be built by means of deep metric learning using paired data consisting of images and their spoken captions. We use it to bridge modality-dependent information, i.e., the speech segments in the mixture, and the specified, modality-independent concept. As a proof of our scheme, we performed experiments using a set of images associated with spoken captions. That is, we generated speech mixtures from these spoken captions and used the images or speech signals as the concept specifiers. We then extracted the target speech using the acoustic characteristics of the identified segments. We compare ConceptBeam with two methods: one based on keywords obtained from recognition systems and another based on sound source separation. We show that ConceptBeam clearly outperforms the baseline methods and effectively extracts speech based on the semantic representation.

Inspired by the recent progress in self-supervised learning for computer vision that generates supervision using data augmentations, we explore a new general-purpose audio representation learning approach. We propose learning general-purpose audio representation from a single audio segment without expecting relationships between different time segments of audio samples. To implement this principle, we introduce Bootstrap Your Own Latent (BYOL) for Audio (BYOL-A, pronounced “viola”), an audio self-supervised learning method based on BYOL for learning general-purpose audio representation. Unlike most previous audio self-supervised learning methods that rely on agreement of vicinity audio segments or disagreement of remote ones, BYOL-A creates contrasts in an augmented audio segment pair derived from a single audio segment. With a combination of normalization and augmentation techniques, BYOL-A achieves state-of-the-art results in various downstream tasks. Extensive ablation studies also clarified the contribution of each component and their combinations.

This paper investigates a self-adaptation method for speech enhancement using auxiliary speaker-aware features; we extract a speaker representation used for adaptation directly from the test utterance. Conventional studies of deep neural network (DNN)-based speech enhancement mainly focus on building a speaker independent model. Meanwhile, in speech applications including speech recognition and synthesis, it is known that model adaptation to the target speaker improves the accuracy. Our research question is whether a DNN for speech enhancement can be adopted to unknown speakers without any auxiliary guidance signal in test-phase. To achieve this, we adopt multi-task learning of speech enhancement and speaker identification, and use the output of the final hidden layer of speaker identification branch as an auxiliary feature. In addition, we use multi-head self-attention for capturing long-term dependencies in the speech and noise. Experimental results on a public dataset show that our strategy achieves the state-of-the-art performance and also outperform conventional methods in terms of subjective quality.

We propose a speech enhancement method using a causal deep neural network (DNN) for real-time applications. DNN has been widely used for estimating a time-frequency (T-F) mask which enhances a speech signal. One popular DNN structure for that is a recurrent neural network (RNN) owing to its capability of effectively modelling time-sequential data like speech. In particular, the long short-term memory (LSTM) is often used to alleviate the vanishing/exploding gradient problem which makes the training of an RNN difficult. However, the number of parameters of LSTM is increased as the price of mitigating the difficulty of training, which requires more computational resources. For real-time speech enhancement, it is preferable to use a smaller network without losing the performance. In this paper, we propose to use the equilibriated recurrent neural network (ERNN) for avoiding the vanishing/exploding gradient problem without increasing the number of parameters. The proposed structure is causal, which requires only the information from the past, in order to apply it in real-time. Compared to the uni- and bi-directional LSTM networks, the proposed method achieved the similar performance with much fewer parameters.

We propose an end-to-end speech enhancement method with trainable time-frequency (T-F) transform based on invertible deep neural network (DNN). The resent development of speech enhancement is brought by using DNN. The ordinary DNN-based speech enhancement employs T-F transform, typically the short-time Fourier transform (STFT), and estimates a T-F mask using DNN. On the other hand, some methods have considered end-to-end networks which directly estimate the enhanced signals without T-F transform. While end-to-end methods have shown promising results, they are black boxes and hard to understand. Therefore, some end-to-end methods used a DNN to learn the linear T-F transform which is much easier to understand. However, the learned transform may not have a property important for ordinary signal processing. In this paper, as the important property of the T-F transform, perfect reconstruction is considered. An invertible nonlinear T-F transform is constructed by DNNs and learned from data so that the obtained transform is perfectly reconstructing filterbank.

We propose a data-driven design method of perfect-reconstruction filterbank (PRFB) for sound-source enhancement (SSE) based on deep neural network (DNN). DNNs have been used to estimate a time-frequency (T-F) mask in the short-time Fourier transform (STFT) domain. Their training is more stable when a simple cost function as mean-squared error (MSE) is utilized comparing to some advanced cost such as objective sound quality assessments. However, such a simple cost function inherits strong assumptions on the statistics of the target and/or noise which is often not satisfied, and the mismatch of assumption results in degraded performance. In this paper, we propose to design the frequency scale of PRFB from training data so that the assumption on MSE is satisfied. For designing the frequency scale, the warped filterbank frame (WFBF) is considered as PRFB. The frequency characteristic of learned WFBF was in between STFT and the wavelet transform, and its effectiveness was confirmed by comparison with a standard STFT-based DNN whose input feature is compressed into the mel scale.

The piano is one of the most popular and attractive musical instruments that leads to a lot of research on it. To synthesize the piano sound in a computer, many modeling methods have been proposed from full physical models to approximated models. The focus of this paper is on the latter, approximating piano sound by an IIR filter. For stably estimating parameters, the Kautz model is chosen as the filter structure. Then, the selection of poles and excitation signal rises as the questions which are typical to the Kautz model that must be solved. In this paper, sparsity based construction of the Kautz model is proposed for approximating piano sound.

Wave-based acoustic simulation methods are studied actively for predicting acoustical phenomena. Finite-difference time-domain (FDTD) method is one of the most popular methods owing to its straightforwardness of calculating an impulse response. In an FDTD simulation, an omnidirectional sound source is usually adopted, which is not realistic because the real sound sources often have specific directivities. However, there is very little research on imposing a directional sound source into FDTD methods. In this paper, a method of realizing a directional sound source in FDTD methods is proposed. It is formulated as an estimation problem of the initial value so that the estimated result corresponds to the desired directivity. The effectiveness of the proposed method is illustrated through some numerical experiments.