Yuma Koizumi

Yuma Koizumi, Heiga Zen, Shigeki Karita et al.

Speech restoration (SR) is a task of converting degraded speech signals into high-quality ones. In this study, we propose a robust SR model called Miipher, and apply Miipher to a new SR application: increasing the amount of high-quality training data for speech generation by converting speech samples collected from the Web to studio-quality. To make our SR model robust against various degradation, we use (i) a speech representation extracted from w2v-BERT for the input feature, and (ii) a text representation extracted from transcripts via PnG-BERT as a linguistic conditioning feature. Experiments show that Miipher (i) is robust against various audio degradation and (ii) enable us to train a high-quality text-to-speech (TTS) model from restored speech samples collected from the Web. Audio samples are available at our demo page: google.github.io/df-conformer/miipher/

Yuma Koizumi, Kohei Yatabe, Heiga Zen et al.

Denoising diffusion probabilistic models (DDPMs) and generative adversarial networks (GANs) are popular generative models for neural vocoders. The DDPMs and GANs can be characterized by the iterative denoising framework and adversarial training, respectively. This study proposes a fast and high-quality neural vocoder called \textit{WaveFit}, which integrates the essence of GANs into a DDPM-like iterative framework based on fixed-point iteration. WaveFit iteratively denoises an input signal, and trains a deep neural network (DNN) for minimizing an adversarial loss calculated from intermediate outputs at all iterations. Subjective (side-by-side) listening tests showed no statistically significant differences in naturalness between human natural speech and those synthesized by WaveFit with five iterations. Furthermore, the inference speed of WaveFit was more than 240 times faster than WaveRNN. Audio demos are available at \url{google.github.io/df-conformer/wavefit/}.

Kota Dohi, Keisuke Imoto, Noboru Harada et al.

We present the task description and discussion on the results of the DCASE 2022 Challenge Task 2: ``Unsupervised anomalous sound detection (ASD) for machine condition monitoring applying domain generalization techniques''. Domain shifts are a critical problem for the application of ASD systems. Because domain shifts can change the acoustic characteristics of data, a model trained in a source domain performs poorly for a target domain. In DCASE 2021 Challenge Task 2, we organized an ASD task for handling domain shifts. In this task, it was assumed that the occurrences of domain shifts are known. However, in practice, the domain of each sample may not be given, and the domain shifts can occur implicitly. In 2022 Task 2, we focus on domain generalization techniques that detects anomalies regardless of the domain shifts. Specifically, the domain of each sample is not given in the test data and only one threshold is allowed for all domains. Analysis of 81 submissions from 31 teams revealed two remarkable types of domain generalization techniques: 1) domain-mixing-based approach that obtains generalized representations and 2) domain-classification-based approach that explicitly or implicitly classifies different domains to improve detection performance for each domain.

Min Ma, Yuma Koizumi, Shigeki Karita et al.

This paper introduces FLEURS-R, a speech restoration applied version of the Few-shot Learning Evaluation of Universal Representations of Speech (FLEURS) corpus. FLEURS-R maintains an N-way parallel speech corpus in 102 languages as FLEURS, with improved audio quality and fidelity by applying the speech restoration model Miipher. The aim of FLEURS-R is to advance speech technology in more languages and catalyze research including text-to-speech (TTS) and other speech generation tasks in low-resource languages. Comprehensive evaluations with the restored speech and TTS baseline models trained from the new corpus show that the new corpus obtained significantly improved speech quality while maintaining the semantic contents of the speech. The corpus is publicly released via Hugging Face.

Gemini Team, Petko Georgiev, Ving Ian Lei et al. · deepmind, mila

In this report, we introduce the Gemini 1.5 family of models, representing the next generation of highly compute-efficient multimodal models capable of recalling and reasoning over fine-grained information from millions of tokens of context, including multiple long documents and hours of video and audio. The family includes two new models: (1) an updated Gemini 1.5 Pro, which exceeds the February version on the great majority of capabilities and benchmarks; (2) Gemini 1.5 Flash, a more lightweight variant designed for efficiency with minimal regression in quality. Gemini 1.5 models achieve near-perfect recall on long-context retrieval tasks across modalities, improve the state-of-the-art in long-document QA, long-video QA and long-context ASR, and match or surpass Gemini 1.0 Ultra's state-of-the-art performance across a broad set of benchmarks. Studying the limits of Gemini 1.5's long-context ability, we find continued improvement in next-token prediction and near-perfect retrieval (>99%) up to at least 10M tokens, a generational leap over existing models such as Claude 3.0 (200k) and GPT-4 Turbo (128k). Finally, we highlight real-world use cases, such as Gemini 1.5 collaborating with professionals on completing their tasks achieving 26 to 75% time savings across 10 different job categories, as well as surprising new capabilities of large language models at the frontier; when given a grammar manual for Kalamang, a language with fewer than 200 speakers worldwide, the model learns to translate English to Kalamang at a similar level to a person who learned from the same content.

Yuma Koizumi, Shoichiro Saito, Hisashi Uematsu et al.

This paper introduces a new dataset called "ToyADMOS" designed for anomaly detection in machine operating sounds (ADMOS). To the best our knowledge, no large-scale datasets are available for ADMOS, although large-scale datasets have contributed to recent advancements in acoustic signal processing. This is because anomalous sound data are difficult to collect. To build a large-scale dataset for ADMOS, we collected anomalous operating sounds of miniature machines (toys) by deliberately damaging them. The released dataset consists of three sub-datasets for machine-condition inspection, fault diagnosis of machines with geometrically fixed tasks, and fault diagnosis of machines with moving tasks. Each sub-dataset includes over 180 hours of normal machine-operating sounds and over 4,000 samples of anomalous sounds collected with four microphones at a 48-kHz sampling rate. The dataset is freely available for download at https://github.com/YumaKoizumi/ToyADMOS-dataset

Shigeki Karita, Yuma Koizumi, Heiga Zen et al.

Training data cleaning is a new application for generative model-based speech restoration (SR). This paper introduces Miipher-2, an SR model designed for million-hour scale data, for training data cleaning for large-scale generative models like large language models. Key challenges addressed include generalization to unseen languages, operation without explicit conditioning (e.g., text, speaker ID), and computational efficiency. Miipher-2 utilizes a frozen, pre-trained Universal Speech Model (USM), supporting over 300 languages, as a robust, conditioning-free feature extractor. To optimize efficiency and minimize memory, Miipher-2 incorporates parallel adapters for predicting clean USM features from noisy inputs and employs the WaveFit neural vocoder for waveform synthesis. These components were trained on 3,000 hours of multi-lingual, studio-quality recordings with augmented degradations, while USM parameters remained fixed. Experimental results demonstrate Miipher-2's superior or comparable performance to conventional SR models in word-error-rate, speaker similarity, and both objective and subjective sound quality scores across all tested languages. Miipher-2 operates efficiently on consumer-grade accelerators, achieving a real-time factor of 0.0078, enabling the processing of a million-hour speech dataset in approximately three days using only 100 such accelerators.

Kota Dohi, Keisuke Imoto, Noboru Harada et al.

We present the task description of the Detection and Classification of Acoustic Scenes and Events (DCASE) 2023 Challenge Task 2: ``First-shot unsupervised anomalous sound detection (ASD) for machine condition monitoring''. The main goal is to enable rapid deployment of ASD systems for new kinds of machines without the need for hyperparameter tuning. In the past ASD tasks, developed methods tuned hyperparameters for each machine type, as the development and evaluation datasets had the same machine types. However, collecting normal and anomalous data as the development dataset can be infeasible in practice. In 2023 Task 2, we focus on solving the first-shot problem, which is the challenge of training a model on a completely novel machine type. Specifically, (i) each machine type has only one section (a subset of machine type) and (ii) machine types in the development and evaluation datasets are completely different. Analysis of 86 submissions from 23 teams revealed that the keys to outperform baselines were: 1) sampling techniques for dealing with class imbalances across different domains and attributes, 2) generation of synthetic samples for robust detection, and 3) use of multiple large pre-trained models to extract meaningful embeddings for the anomaly detector.

Yuma Koizumi, Heiga Zen, Kohei Yatabe et al.

Neural vocoder using denoising diffusion probabilistic model (DDPM) has been improved by adaptation of the diffusion noise distribution to given acoustic features. In this study, we propose SpecGrad that adapts the diffusion noise so that its time-varying spectral envelope becomes close to the conditioning log-mel spectrogram. This adaptation by time-varying filtering improves the sound quality especially in the high-frequency bands. It is processed in the time-frequency domain to keep the computational cost almost the same as the conventional DDPM-based neural vocoders. Experimental results showed that SpecGrad generates higher-fidelity speech waveform than conventional DDPM-based neural vocoders in both analysis-synthesis and speech enhancement scenarios. Audio demos are available at wavegrad.github.io/specgrad/.

Yuma Koizumi, Shigeki Karita, Arun Narayanan et al.

Speech enhancement (SE) is used as a frontend in speech applications including automatic speech recognition (ASR) and telecommunication. A difficulty in using the SE frontend is that the appropriate noise reduction level differs depending on applications and/or noise characteristics. In this study, we propose "signal-to-noise ratio improvement (SNRi) target training"; the SE frontend is trained to output a signal whose SNRi is controlled by an auxiliary scalar input. In joint training with a backend, the target SNRi value is estimated by an auxiliary network. By training all networks to minimize the backend task loss, we can estimate the appropriate noise reduction level for each noisy input in a data-driven scheme. Our experiments showed that the SNRi target training enables control of the output SNRi. In addition, the proposed joint training relatively reduces word error rate by 4.0\% and 5.7\% compared to a Conformer-based standard ASR model and conventional SE-ASR joint training model, respectively. Furthermore, by analyzing the predicted target SNRi, we observed the jointly trained network automatically controls the target SNRi according to noise characteristics. Audio demos are available in our demo page: google.github.io/df-conformer/snri_target/.

Yuma Koizumi, Shigeki Karita, Scott Wisdom et al.

Single-channel speech enhancement (SE) is an important task in speech processing. A widely used framework combines an analysis/synthesis filterbank with a mask prediction network, such as the Conv-TasNet architecture. In such systems, the denoising performance and computational efficiency are mainly affected by the structure of the mask prediction network. In this study, we aim to improve the sequential modeling ability of Conv-TasNet architectures by integrating Conformer layers into a new mask prediction network. To make the model computationally feasible, we extend the Conformer using linear complexity attention and stacked 1-D dilated depthwise convolution layers. We trained the model on 3,396 hours of noisy speech data, and show that (i) the use of linear complexity attention avoids high computational complexity, and (ii) our model achieves higher scale-invariant signal-to-noise ratio than the improved time-dilated convolution network (TDCN++), an extended version of Conv-TasNet.

Yohei Kawaguchi, Keisuke Imoto, Yuma Koizumi et al.

We present the task description and discussion on the results of the DCASE 2021 Challenge Task 2. In 2020, we organized an unsupervised anomalous sound detection (ASD) task, identifying whether a given sound was normal or anomalous without anomalous training data. In 2021, we organized an advanced unsupervised ASD task under domain-shift conditions, which focuses on the inevitable problem of the practical use of ASD systems. The main challenge of this task is to detect unknown anomalous sounds where the acoustic characteristics of the training and testing samples are different, i.e., domain-shifted. This problem frequently occurs due to changes in seasons, manufactured products, and/or environmental noise. We received 75 submissions from 26 teams, and several novel approaches have been developed in this challenge. On the basis of the analysis of the evaluation results, we found that there are two types of remarkable approaches that TOP-5 winning teams adopted: 1) ensemble approaches of ``outlier exposure'' (OE)-based detectors and ``inlier modeling'' (IM)-based detectors and 2) approaches based on IM-based detection for features learned in a machine-identification task.

Koichi Saito, Tomohiko Nakamura, Kohei Yatabe et al.

Audio source separation is often used as preprocessing of various applications, and one of its ultimate goals is to construct a single versatile model capable of dealing with the varieties of audio signals. Since sampling frequency, one of the audio signal varieties, is usually application specific, the preceding audio source separation model should be able to deal with audio signals of all sampling frequencies specified in the target applications. However, conventional models based on deep neural networks (DNNs) are trained only at the sampling frequency specified by the training data, and there are no guarantees that they work with unseen sampling frequencies. In this paper, we propose a convolution layer capable of handling arbitrary sampling frequencies by a single DNN. Through music source separation experiments, we show that the introduction of the proposed layer enables a conventional audio source separation model to consistently work with even unseen sampling frequencies.

Yuma Koizumi, Yasunori Ohishi, Daisuke Niizumi et al.

The goal of audio captioning is to translate input audio into its description using natural language. One of the problems in audio captioning is the lack of training data due to the difficulty in collecting audio-caption pairs by crawling the web. In this study, to overcome this problem, we propose to use a pre-trained large-scale language model. Since an audio input cannot be directly inputted into such a language model, we utilize guidance captions retrieved from a training dataset based on similarities that may exist in different audio. Then, the caption of the audio input is generated by using a pre-trained language model while referring to the guidance captions. Experimental results show that (i) the proposed method has succeeded to use a pre-trained language model for audio captioning, and (ii) the oracle performance of the pre-trained model-based caption generator was clearly better than that of the conventional method trained from scratch.

Daiki Takeuchi, Yuma Koizumi, Yasunori Ohishi et al.

The system we used for Task 6 (Automated Audio Captioning)of the Detection and Classification of Acoustic Scenes and Events(DCASE) 2020 Challenge combines three elements, namely, dataaugmentation, multi-task learning, and post-processing, for audiocaptioning. The system received the highest evaluation scores, butwhich of the individual elements most fully contributed to its perfor-mance has not yet been clarified. Here, to asses their contributions,we first conducted an element-wise ablation study on our systemto estimate to what extent each element is effective. We then con-ducted a detailed module-wise ablation study to further clarify thekey processing modules for improving accuracy. The results showthat data augmentation and post-processing significantly improvethe score in our system. In particular, mix-up data augmentationand beam search in post-processing improve SPIDEr by 0.8 and 1.6points, respectively.

Yuma Koizumi, Daiki Takeuchi, Yasunori Ohishi et al.

This technical report describes the system participating to the Detection and Classification of Acoustic Scenes and Events (DCASE) 2020 Challenge, Task 6: automated audio captioning. Our submission focuses on solving two indeterminacy problems in automated audio captioning: word selection indeterminacy and sentence length indeterminacy. We simultaneously solve the main caption generation and sub indeterminacy problems by estimating keywords and sentence length through multi-task learning. We tested a simplified model of our submission using the development-testing dataset. Our model achieved 20.7 SPIDEr score where that of the baseline system was 5.4.

Yuma Koizumi, Ryo Masumura, Kyosuke Nishida et al.

One of the problems with automated audio captioning (AAC) is the indeterminacy in word selection corresponding to the audio event/scene. Since one acoustic event/scene can be described with several words, it results in a combinatorial explosion of possible captions and difficulty in training. To solve this problem, we propose a Transformer-based audio-captioning model with keyword estimation called TRACKE. It simultaneously solves the word-selection indeterminacy problem with the main task of AAC while executing the sub-task of acoustic event detection/acoustic scene classification (i.e., keyword estimation). TRACKE estimates keywords, which comprise a word set corresponding to audio events/scenes in the input audio, and generates the caption while referring to the estimated keywords to reduce word-selection indeterminacy. Experimental results on a public AAC dataset indicate that TRACKE achieved state-of-the-art performance and successfully estimated both the caption and its keywords.

Yuma Koizumi, Yohei Kawaguchi, Keisuke Imoto et al.

In this paper, we present the task description and discuss the results of the DCASE 2020 Challenge Task 2: Unsupervised Detection of Anomalous Sounds for Machine Condition Monitoring. The goal of anomalous sound detection (ASD) is to identify whether the sound emitted from a target machine is normal or anomalous. The main challenge of this task is to detect unknown anomalous sounds under the condition that only normal sound samples have been provided as training data. We have designed this challenge as the first benchmark of ASD research, which includes a large-scale dataset, evaluation metrics, and a simple baseline system. We received 117 submissions from 40 teams, and several novel approaches have been developed as a result of this challenge. On the basis of the analysis of the evaluation results, we discuss two new approaches and their problems.

Tsubasa Ochiai, Marc Delcroix, Yuma Koizumi et al.

Being able to control the acoustic events (AEs) to which we want to listen would allow the development of more controllable hearable devices. This paper addresses the AE sound selection (or removal) problems, that we define as the extraction (or suppression) of all the sounds that belong to one or multiple desired AE classes. Although this problem could be addressed with a combination of source separation followed by AE classification, this is a sub-optimal way of solving the problem. Moreover, source separation usually requires knowing the maximum number of sources, which may not be practical when dealing with AEs. In this paper, we propose instead a universal sound selection neural network that enables to directly select AE sounds from a mixture given user-specified target AE classes. The proposed framework can be explicitly optimized to simultaneously select sounds from multiple desired AE classes, independently of the number of sources in the mixture. We experimentally show that the proposed method achieves promising AE sound selection performance and could be generalized to mixtures with a number of sources that are unseen during training.

Masahiro Yasuda, Yuma Koizumi, Shoichiro Saito et al.

We propose a direction-of-arrival (DOA) estimation method for Sound Event Localization and Detection (SELD). Direct estimation of DOA using a deep neural network (DNN), i.e. completely-datadriven approach, achieves high accuracy. However, there is a gap in the accuracy between DOA estimation for single and overlapping sources because they cannot incorporate physical knowledge. Meanwhile, although the accuracy of physics-based approaches is inferior to DNN-based approaches, it is robust for overlapping source. In this study, we consider a combination of physics-based and DNN-based approaches; the sound intensity vectors (IVs) for physics-based DOA estimation is refined based on DNN-based denoising and source separation. This method enables the accurate DOA estimation for both single and overlapping sources using a spherical microphone array. Experimental results show that the proposed method achieves state-of-the-art DOA estimation accuracy on an open dataset of the SELD.

Masaki Kawanaka, Yuma Koizumi, Ryoichi Miyazaki et al.

Improving subjective sound quality of enhanced signals is one of the most important missions in speech enhancement. For evaluating the subjective quality, several methods related to perceptually-motivated objective sound quality assessment (OSQA) have been proposed such as PESQ (perceptual evaluation of speech quality). However, direct use of such measures for training deep neural network (DNN) is not allowed in most cases because popular OSQAs are non-differentiable with respect to DNN parameters. Therefore, the previous study has proposed to approximate the score of OSQAs by an auxiliary DNN so that its gradient can be used for training the primary DNN. One problem with this approach is instability of the training caused by the approximation error of the score. To overcome this problem, we propose to use stabilization techniques borrowed from reinforcement learning. The experiments, aimed to increase the score of PESQ as an example, show that the proposed method (i) can stably train a DNN to increase PESQ, (ii) achieved the state-of-the-art PESQ score on a public dataset, and (iii) resulted in better sound quality than conventional methods based on subjective evaluation.

Yuma Koizumi, Kohei Yatabe, Marc Delcroix et al.

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.

Keisuke Imoto, Noriyuki Tonami, Yuma Koizumi et al.

Sound event detection (SED) and acoustic scene classification (ASC) are major tasks in environmental sound analysis. Considering that sound events and scenes are closely related to each other, some works have addressed joint analyses of sound events and acoustic scenes based on multitask learning (MTL), in which the knowledge of sound events and scenes can help in estimating them mutually. The conventional MTL-based methods utilize one-hot scene labels to train the relationship between sound events and scenes; thus, the conventional methods cannot model the extent to which sound events and scenes are related. However, in the real environment, common sound events may occur in some acoustic scenes; on the other hand, some sound events occur only in a limited acoustic scene. In this paper, we thus propose a new method for SED based on MTL of SED and ASC using the soft labels of acoustic scenes, which enable us to model the extent to which sound events and scenes are related. Experiments conducted using TUT Sound Events 2016/2017 and TUT Acoustic Scenes 2016 datasets show that the proposed method improves the SED performance by 3.80% in F-score compared with conventional MTL-based SED.

Daiki Takeuchi, Kohei Yatabe, Yuma Koizumi et al.

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.

Yoshiki Masuyama, Kohei Yatabe, Yuma Koizumi et al.

Phase reconstruction, which estimates phase from a given amplitude spectrogram, is an active research field in acoustical signal processing with many applications including audio synthesis. To take advantage of rich knowledge from data, several studies presented deep neural network (DNN)--based phase reconstruction methods. However, the training of a DNN for phase reconstruction is not an easy task because phase is sensitive to the shift of a waveform. To overcome this problem, we propose a DNN-based two-stage phase reconstruction method. In the proposed method, DNNs estimate phase derivatives instead of phase itself, which allows us to avoid the sensitivity problem. Then, phase is recursively estimated based on the estimated derivatives, which is named recurrent phase unwrapping (RPU). The experimental results confirm that the proposed method outperformed the direct phase estimation by a DNN.

Daiki Takeuchi, Kohei Yatabe, Yuma Koizumi et al.

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.

Masahiro Yasuda, Yuma Koizumi, Luca Mazzon et al.

We propose a direction of arrival (DOA) estimation method that combines sound-intensity vector (IV)-based DOA estimation and DNN-based denoising and dereverberation. Since the accuracy of IV-based DOA estimation degrades due to environmental noise and reverberation, two DNNs are used to remove such effects from the observed IVs. DOA is then estimated from the refined IVs based on the physics of wave propagation. Experiments on an open dataset showed that the average DOA error of the proposed method was 0.528 degrees, and it outperformed a conventional IV-based and DNN-based DOA estimation method.

Luca Mazzon, Yuma Koizumi, Masahiro Yasuda et al.

In this paper, we propose a novel data augmentation method for training neural networks for Direction of Arrival (DOA) estimation. This method focuses on expanding the representation of the DOA subspace of a dataset. Given some input data, it applies a transformation to it in order to change its DOA information and simulate new potentially unseen one. Such transformation, in general, is a combination of a rotation and a reflection. It is possible to apply such transformation due to a well-known property of First Order Ambisonics (FOA). The same transformation is applied also to the labels, in order to maintain consistency between input data and target labels. Three methods with different level of generality are proposed for applying this augmentation principle. Experiments are conducted on two different DOA networks. Results of both experiments demonstrate the effectiveness of the novel augmentation strategy by improving the DOA error by around 40%.

Yuma Koizumi, Shoichiro Saito, Masataka Yamaguchi et al.

Use of an autoencoder (AE) as a normal model is a state-of-the-art technique for unsupervised-anomaly detection in sounds (ADS). The AE is trained to minimize the sample mean of the anomaly score of normal sounds in a mini-batch. One problem with this approach is that the anomaly score of rare-normal sounds becomes higher than that of frequent-normal sounds, because the sample mean is strongly affected by frequent-normal samples, resulting in preferentially decreasing the anomaly score of frequent-normal samples. To decrease anomaly scores for both frequent- and rare-normal sounds, we propose batch uniformization, a training method for unsupervised-ADS for minimizing a weighted average of the anomaly score on each sample in a mini-batch. We used the reciprocal of the probabilistic density of each sample as the weight, more intuitively, a large weight is given for rare-normal sounds. Such a weight works to give a constant anomaly score for both frequent- and rare-normal sounds. Since the probabilistic density is unknown, we estimate it by using the kernel density estimation on each training mini-batch. Verification- and objective-experiments show that the proposed batch uniformization improves the performance of unsupervised-ADS.

Daiki Takeuchi, Kohei Yatabe, Yuma Koizumi et al.

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.

Yoshiki Masuyama, Kohei Yatabe, Yuma Koizumi et al.

This paper presents a novel phase reconstruction method (only from a given amplitude spectrogram) by combining a signal-processing-based approach and a deep neural network (DNN). To retrieve a time-domain signal from its amplitude spectrogram, the corresponding phase is required. One of the popular phase reconstruction methods is the Griffin-Lim algorithm (GLA), which is based on the redundancy of the short-time Fourier transform. However, GLA often involves many iterations and produces low-quality signals owing to the lack of prior knowledge of the target signal. In order to address these issues, in this study, we propose an architecture which stacks a sub-block including two GLA-inspired fixed layers and a DNN. The number of stacked sub-blocks is adjustable, and we can trade the performance and computational load based on requirements of applications. The effectiveness of the proposed method is investigated by reconstructing phases from amplitude spectrograms of speeches.

Masataka Yamaguchi, Yuma Koizumi, Noboru Harada

We tackle unsupervised anomaly detection (UAD), a problem of detecting data that significantly differ from normal data. UAD is typically solved by using density estimation. Recently, deep neural network (DNN)-based density estimators, such as Normalizing Flows, have been attracting attention. However, one of their drawbacks is the difficulty in adapting them to the change in the normal data's distribution. To address this difficulty, we propose AdaFlow, a new DNN-based density estimator that can be easily adapted to the change of the distribution. AdaFlow is a unified model of a Normalizing Flow and Adaptive Batch-Normalizations, a module that enables DNNs to adapt to new distributions. AdaFlow can be adapted to a new distribution by just conducting forward propagation once per sample; hence, it can be used on devices that have limited computational resources. We have confirmed the effectiveness of the proposed model through an anomaly detection in a sound task. We also propose a method of applying AdaFlow to the unpaired cross-domain translation problem, in which one has to train a cross-domain translation model with only unpaired samples. We have confirmed that our model can be used for the cross-domain translation problem through experiments on image datasets.

Yuma Koizumi, Noboru Harada, Yoichi Haneda

This study proposes a trainable adaptive window switching (AWS) method and apply it to a deep-neural-network (DNN) for speech enhancement in the modified discrete cosine transform domain. Time-frequency (T-F) mask processing in the short-time Fourier transform (STFT)-domain is a typical speech enhancement method. To recover the target signal precisely, DNN-based short-time frequency transforms have recently been investigated and used instead of the STFT. However, since such a fixed-resolution short-time frequency transform method has a T-F resolution problem based on the uncertainty principle, not only the short-time frequency transform but also the length of the windowing function should be optimized. To overcome this problem, we incorporate AWS into the speech enhancement procedure, and the windowing function of each time-frame is manipulated using a DNN depending on the input signal. We confirmed that the proposed method achieved a higher signal-to-distortion ratio than conventional speech enhancement methods in fixed-resolution frequency domains.

Yuma Koizumi, Kenta Niwa, Yusuke Hioka et al.

We propose a training method for deep neural network (DNN)-based source enhancement to increase objective sound quality assessment (OSQA) scores such as the perceptual evaluation of speech quality (PESQ). In many conventional studies, DNNs have been used as a mapping function to estimate time-frequency masks and trained to minimize an analytically tractable objective function such as the mean squared error (MSE). Since OSQA scores have been used widely for sound-quality evaluation, constructing DNNs to increase OSQA scores would be better than using the minimum-MSE to create high-quality output signals. However, since most OSQA scores are not analytically tractable, \textit{i.e.}, they are black boxes, the gradient of the objective function cannot be calculated by simply applying back-propagation. To calculate the gradient of the OSQA-based objective function, we formulated a DNN optimization scheme on the basis of \textit{black-box optimization}, which is used for training a computer that plays a game. For a black-box-optimization scheme, we adopt the policy gradient method for calculating the gradient on the basis of a sampling algorithm. To simulate output signals using the sampling algorithm, DNNs are used to estimate the probability density function of the output signals that maximize OSQA scores. The OSQA scores are calculated from the simulated output signals, and the DNNs are trained to increase the probability of generating the simulated output signals that achieve high OSQA scores. Through several experiments, we found that OSQA scores significantly increased by applying the proposed method, even though the MSE was not minimized.

Yuma Koizumi, Shoichiro Saito, Hisashi Uematsum Yuta Kawachi et al.

This paper proposes a novel optimization principle and its implementation for unsupervised anomaly detection in sound (ADS) using an autoencoder (AE). The goal of unsupervised-ADS is to detect unknown anomalous sound without training data of anomalous sound. Use of an AE as a normal model is a state-of-the-art technique for unsupervised-ADS. To decrease the false positive rate (FPR), the AE is trained to minimize the reconstruction error of normal sounds and the anomaly score is calculated as the reconstruction error of the observed sound. Unfortunately, since this training procedure does not take into account the anomaly score for anomalous sounds, the true positive rate (TPR) does not necessarily increase. In this study, we define an objective function based on the Neyman-Pearson lemma by considering ADS as a statistical hypothesis test. The proposed objective function trains the AE to maximize the TPR under an arbitrary low FPR condition. To calculate the TPR in the objective function, we consider that the set of anomalous sounds is the complementary set of normal sounds and simulate anomalous sounds by using a rejection sampling algorithm. Through experiments using synthetic data, we found that the proposed method improved the performance measures of ADS under low FPR conditions. In addition, we confirmed that the proposed method could detect anomalous sounds in real environments.