Hiroshi Saruwatari

Wataru Nakata, Kentaro Seki, Hitomi Yanaka et al.

Spoken dialogue plays a crucial role in human-AI interactions, necessitating dialogue-oriented spoken language models (SLMs). To develop versatile SLMs, large-scale and diverse speech datasets are essential. Additionally, to ensure hiqh-quality speech generation, the data must be spontaneous like in-wild data and must be acoustically clean with noise removed. Despite the critical need, no open-source corpus meeting all these criteria has been available. This study addresses this gap by constructing and releasing a large-scale spoken dialogue corpus, named Japanese Corpus for Human-AI Talks (J-CHAT), which is publicly accessible. Furthermore, this paper presents a language-independent method for corpus construction and describes experiments on dialogue generation using SLMs trained on J-CHAT. Experimental results indicate that the collected data from multiple domains by our method improve the naturalness and meaningfulness of dialogue generation.

Takaaki Saeki, Soumi Maiti, Xinjian Li et al.

While neural text-to-speech (TTS) has achieved human-like natural synthetic speech, multilingual TTS systems are limited to resource-rich languages due to the need for paired text and studio-quality audio data. This paper proposes a method for zero-shot multilingual TTS using text-only data for the target language. The use of text-only data allows the development of TTS systems for low-resource languages for which only textual resources are available, making TTS accessible to thousands of languages. Inspired by the strong cross-lingual transferability of multilingual language models, our framework first performs masked language model pretraining with multilingual text-only data. Then we train this model with a paired data in a supervised manner, while freezing a language-aware embedding layer. This allows inference even for languages not included in the paired data but present in the text-only data. Evaluation results demonstrate highly intelligible zero-shot TTS with a character error rate of less than 12% for an unseen language.

Yuki Saito, Yuto Nishimura, Shinnosuke Takamichi et al.

We present STUDIES, a new speech corpus for developing a voice agent that can speak in a friendly manner. Humans naturally control their speech prosody to empathize with each other. By incorporating this "empathetic dialogue" behavior into a spoken dialogue system, we can develop a voice agent that can respond to a user more naturally. We designed the STUDIES corpus to include a speaker who speaks with empathy for the interlocutor's emotion explicitly. We describe our methodology to construct an empathetic dialogue speech corpus and report the analysis results of the STUDIES corpus. We conducted a text-to-speech experiment to initially investigate how we can develop more natural voice agent that can tune its speaking style corresponding to the interlocutor's emotion. The results show that the use of interlocutor's emotion label and conversational context embedding can produce speech with the same degree of naturalness as that synthesized by using the agent's emotion label. Our project page of the STUDIES corpus is http://sython.org/Corpus/STUDIES.

Tomohiko Nakamura, Shinnosuke Takamichi, Naoko Tanji et al.

We construct a corpus of Japanese a cappella vocal ensembles (jaCappella corpus) for vocal ensemble separation and synthesis. It consists of 35 copyright-cleared vocal ensemble songs and their audio recordings of individual voice parts. These songs were arranged from out-of-copyright Japanese children's songs and have six voice parts (lead vocal, soprano, alto, tenor, bass, and vocal percussion). They are divided into seven subsets, each of which features typical characteristics of a music genre such as jazz and enka. The variety in genre and voice part match vocal ensembles recently widespread in social media services such as YouTube, although the main targets of conventional vocal ensemble datasets are choral singing made up of soprano, alto, tenor, and bass. Experimental evaluation demonstrates that our corpus is a challenging resource for vocal ensemble separation. Our corpus is available on our project page (https://tomohikonakamura.github.io/jaCappella_corpus/).

Dong Yang, Tomoki Koriyama, Yuki Saito et al.

Pause insertion, also known as phrase break prediction and phrasing, is an essential part of TTS systems because proper pauses with natural duration significantly enhance the rhythm and intelligibility of synthetic speech. However, conventional phrasing models ignore various speakers' different styles of inserting silent pauses, which can degrade the performance of the model trained on a multi-speaker speech corpus. To this end, we propose more powerful pause insertion frameworks based on a pre-trained language model. Our approach uses bidirectional encoder representations from transformers (BERT) pre-trained on a large-scale text corpus, injecting speaker embedding to capture various speaker characteristics. We also leverage duration-aware pause insertion for more natural multi-speaker TTS. We develop and evaluate two types of models. The first improves conventional phrasing models on the position prediction of respiratory pauses (RPs), i.e., silent pauses at word transitions without punctuation. It performs speaker-conditioned RP prediction considering contextual information and is used to demonstrate the effect of speaker information on the prediction. The second model is further designed for phoneme-based TTS models and performs duration-aware pause insertion, predicting both RPs and punctuation-indicated pauses (PIPs) that are categorized by duration. The evaluation results show that our models improve the precision and recall of pause insertion and the rhythm of synthetic speech.

Yuto Nishimura, Yuki Saito, Shinnosuke Takamichi et al.

We propose an end-to-end empathetic dialogue speech synthesis (DSS) model that considers both the linguistic and prosodic contexts of dialogue history. Empathy is the active attempt by humans to get inside the interlocutor in dialogue, and empathetic DSS is a technology to implement this act in spoken dialogue systems. Our model is conditioned by the history of linguistic and prosody features for predicting appropriate dialogue context. As such, it can be regarded as an extension of the conventional linguistic-feature-based dialogue history modeling. To train the empathetic DSS model effectively, we investigate 1) a self-supervised learning model pretrained with large speech corpora, 2) a style-guided training using a prosody embedding of the current utterance to be predicted by the dialogue context embedding, 3) a cross-modal attention to combine text and speech modalities, and 4) a sentence-wise embedding to achieve fine-grained prosody modeling rather than utterance-wise modeling. The evaluation results demonstrate that 1) simply considering prosodic contexts of the dialogue history does not improve the quality of speech in empathetic DSS and 2) introducing style-guided training and sentence-wise embedding modeling achieves higher speech quality than that by the conventional method.

Shinnosuke Takamichi, Hiroki Maeda, Joonyong Park et al.

In this study, we investigate whether speech symbols, learned through deep learning, follow Zipf's law, akin to natural language symbols. Zipf's law is an empirical law that delineates the frequency distribution of words, forming fundamentals for statistical analysis in natural language processing. Natural language symbols, which are invented by humans to symbolize speech content, are recognized to comply with this law. On the other hand, recent breakthroughs in spoken language processing have given rise to the development of learned speech symbols; these are data-driven symbolizations of speech content. Our objective is to ascertain whether these data-driven speech symbols follow Zipf's law, as the same as natural language symbols. Through our investigation, we aim to forge new ways for the statistical analysis of spoken language processing.

Joonyong Park, Shinnosuke Takamichi, Tomohiko Nakamura et al.

We examine the speech modeling potential of generative spoken language modeling (GSLM), which involves using learned symbols derived from data rather than phonemes for speech analysis and synthesis. Since GSLM facilitates textless spoken language processing, exploring its effectiveness is critical for paving the way for novel paradigms in spoken-language processing. This paper presents the findings of GSLM's encoding and decoding effectiveness at the spoken-language and speech levels. Through speech resynthesis experiments, we revealed that resynthesis errors occur at the levels ranging from phonology to syntactics and GSLM frequently resynthesizes natural but content-altered speech.

Yuta Matsunaga, Takaaki Saeki, Shinnosuke Takamichi et al.

We present a comprehensive empirical study for personalized spontaneous speech synthesis on the basis of linguistic knowledge. With the advent of voice cloning for reading-style speech synthesis, a new voice cloning paradigm for human-like and spontaneous speech synthesis is required. We, therefore, focus on personalized spontaneous speech synthesis that can clone both the individual's voice timbre and speech disfluency. Specifically, we deal with filled pauses, a major source of speech disfluency, which is known to play an important role in speech generation and communication in psychology and linguistics. To comparatively evaluate personalized filled pause insertion and non-personalized filled pause prediction methods, we developed a speech synthesis method with a non-personalized external filled pause predictor trained with a multi-speaker corpus. The results clarify the position-word entanglement of filled pauses, i.e., the necessity of precisely predicting positions for naturalness and the necessity of precisely predicting words for individuality on the evaluation of synthesized speech.

Yusuke Nakai, Yuki Saito, Kenta Udagawa et al.

We propose a novel training algorithm for a multi-speaker neural text-to-speech (TTS) model based on multi-task adversarial training. A conventional generative adversarial network (GAN)-based training algorithm significantly improves the quality of synthetic speech by reducing the statistical difference between natural and synthetic speech. However, the algorithm does not guarantee the generalization performance of the trained TTS model in synthesizing voices of unseen speakers who are not included in the training data. Our algorithm alternatively trains two deep neural networks: multi-task discriminator and multi-speaker neural TTS model (i.e., generator of GANs). The discriminator is trained not only to distinguish between natural and synthetic speech but also to verify the speaker of input speech is existent or non-existent (i.e., newly generated by interpolating seen speakers' embedding vectors). Meanwhile, the generator is trained to minimize the weighted sum of the speech reconstruction loss and adversarial loss for fooling the discriminator, which achieves high-quality multi-speaker TTS even if the target speaker is unseen. Experimental evaluation shows that our algorithm improves the quality of synthetic speech better than a conventional GANSpeech algorithm.

Kaito Baba, Wataru Nakata, Yuki Saito et al.

We present our system (denoted as T05) for the VoiceMOS Challenge (VMC) 2024. Our system was designed for the VMC 2024 Track 1, which focused on the accurate prediction of naturalness mean opinion score (MOS) for high-quality synthetic speech. In addition to a pretrained self-supervised learning (SSL)-based speech feature extractor, our system incorporates a pretrained image feature extractor to capture the difference of synthetic speech observed in speech spectrograms. We first separately train two MOS predictors that use either of an SSL-based or spectrogram-based feature. Then, we fine-tune the two predictors for better MOS prediction using the fusion of two extracted features. In the VMC 2024 Track 1, our T05 system achieved first place in 7 out of 16 evaluation metrics and second place in the remaining 9 metrics, with a significant difference compared to those ranked third and below. We also report the results of our ablation study to investigate essential factors of our system.

Kazuki Yamauchi, Yuki Saito, Hiroshi Saruwatari

We explore cross-dialect text-to-speech (CD-TTS), a task to synthesize learned speakers' voices in non-native dialects, especially in pitch-accent languages. CD-TTS is important for developing voice agents that naturally communicate with people across regions. We present a novel TTS model comprising three sub-modules to perform competitively at this task. We first train a backbone TTS model to synthesize dialect speech from a text conditioned on phoneme-level accent latent variables (ALVs) extracted from speech by a reference encoder. Then, we train an ALV predictor to predict ALVs tailored to a target dialect from input text leveraging our novel multi-dialect phoneme-level BERT. We conduct multi-dialect TTS experiments and evaluate the effectiveness of our model by comparing it with a baseline derived from conventional dialect TTS methods. The results show that our model improves the dialectal naturalness of synthetic speech in CD-TTS.

Dong Yang, Yiyi Cai, Haoyu Zhang et al.

Metric-induced discrete flow matching (MI-DFM) exploits token-latent geometry for discrete generation, but its practical use is limited by two issues: heuristic schedulers requiring hyperparameter search, and finite-step path-tracking error from its first-order continuous-time Markov chain (CTMC) solver. We address both issues. First, we derive a kinetic-optimal scheduler for prescribed scalar-parameterized probability paths, and instantiate it for MI-DFM as a training-free numerical schedule that traverses the path at constant Fisher-Rao speed. Second, we introduce a finite-step moment correction that adjusts the jump probability while preserving the CTMC jump destination distribution. We validate the resulting method, GibbsTTS, on codec-based zero-shot text-to-speech (TTS). Under controlled comparisons with a unified architecture and large-scale dataset, GibbsTTS achieves the best objective naturalness and is preferred in subjective evaluations over masked discrete generative baselines. Additionally, in comparison with the evaluated state-of-the-art TTS systems, GibbsTTS shows strong speaker similarity, achieving the highest similarity on three of four test sets and ranking second on the fourth. Project page: https://ydqmkkx.github.io/GibbsTTSProject

Shinnosuke Takamichi, Wataru Nakata, Naoko Tanji et al.

In this paper, we construct a Japanese audiobook speech corpus called "J-MAC" for speech synthesis research. With the success of reading-style speech synthesis, the research target is shifting to tasks that use complicated contexts. Audiobook speech synthesis is a good example that requires cross-sentence, expressiveness, etc. Unlike reading-style speech, speaker-specific expressiveness in audiobook speech also becomes the context. To enhance this research, we propose a method of constructing a corpus from audiobooks read by professional speakers. From many audiobooks and their texts, our method can automatically extract and refine the data without any language dependency. Specifically, we use vocal-instrumental separation to extract clean data, connectionist temporal classification to roughly align text and audio, and voice activity detection to refine the alignment. J-MAC is open-sourced in our project page. We also conduct audiobook speech synthesis evaluations, and the results give insights into audiobook speech synthesis.

Detai Xin, Xu Tan, Kai Shen et al.

We present RALL-E, a robust language modeling method for text-to-speech (TTS) synthesis. While previous work based on large language models (LLMs) shows impressive performance on zero-shot TTS, such methods often suffer from poor robustness, such as unstable prosody (weird pitch and rhythm/duration) and a high word error rate (WER), due to the autoregressive prediction style of language models. The core idea behind RALL-E is chain-of-thought (CoT) prompting, which decomposes the task into simpler steps to enhance the robustness of LLM-based TTS. To accomplish this idea, RALL-E first predicts prosody features (pitch and duration) of the input text and uses them as intermediate conditions to predict speech tokens in a CoT style. Second, RALL-E utilizes the predicted duration prompt to guide the computing of self-attention weights in Transformer to enforce the model to focus on the corresponding phonemes and prosody features when predicting speech tokens. Results of comprehensive objective and subjective evaluations demonstrate that, compared to a powerful baseline method VALL-E, RALL-E significantly improves the WER of zero-shot TTS from $5.6\%$ (without reranking) and $1.7\%$ (with reranking) to $2.5\%$ and $1.0\%$, respectively. Furthermore, we demonstrate that RALL-E correctly synthesizes sentences that are hard for VALL-E and reduces the error rate from $68\%$ to $4\%$.

Wataru Nakata, Yuki Saito, Kazuki Yamauchi et al.

Full-duplex dialogue audio, in which each speaker is recorded on a separate track, is an important resource for spoken dialogue research, but is difficult to collect at scale. Most in-the-wild two-speaker dialogue is available only as degraded monaural mixtures, making it unsuitable for systems requiring clean speaker-wise signals. We propose DialogueSidon, a model for joint restoration and separation of degraded monaural two-speaker dialogue audio. DialogueSidon combines a variational autoencoder (VAE) operates on the speech self-supervised learning (SSL) model feature, which compresses SSL model features into a compact latent space, with a diffusion-based latent predictor that recovers speaker-wise latent representations from the degraded mixture. Experiments on English, multilingual, and in-the-wild dialogue datasets show that DialogueSidon substantially improves intelligibility and separation quality over a baseline, while also achieving much faster inference.

Jianing Yang, Sheng Li, Takahiro Shinozaki et al.

Current emotional Text-To-Speech (TTS) and style transfer methods rely on reference encoders to control global style or emotion vectors, but do not capture nuanced acoustic details of the reference speech. To this end, we propose a novel emotional TTS method that enables fine-grained phoneme-level emotion embedding prediction while disentangling intrinsic attributes of the reference speech. The proposed method employs a style disentanglement method to guide two feature extractors, reducing mutual information between timbre and emotion features, and effectively separating distinct style components from the reference speech. Experimental results demonstrate that our method outperforms baseline TTS systems in generating natural and emotionally rich speech. This work highlights the potential of disentangled and fine-grained representations in advancing the quality and flexibility of emotional TTS systems.

Joonyong Park, Shinnosuke Takamichi, David M. Chan et al.

This study presents a comparative analysis of the statistical and linguistic properties of neural audio codecs (NACs). We investigate discrete speech tokens produced by various NAC models, examining their adherence to linguistic statistical laws such as Zipf's law and Heaps' law, as well as their entropy and redundancy. To assess how these token-level properties relate to semantic and acoustic preservation in synthesized speech, we evaluate intelligibility using error rates of automatic speech recognition, and quality using the UTMOS score. Our results reveal that NAC tokens, particularly 3-grams, exhibit language-like statistical patterns. Moreover, these properties, together with measures of information content, are found to correlate with improved performances in speech recognition and resynthesis tasks. These findings offer insights into the structure of NAC token sequences and inform the design of more effective generative speech models.

Dong Yang, Yiyi Cai, Yuki Saito et al.

We propose Shallow Flow Matching (SFM), a novel mechanism that enhances flow matching (FM)-based text-to-speech (TTS) models within a coarse-to-fine generation paradigm. Unlike conventional FM modules, which use the coarse representations from the weak generator as conditions, SFM constructs intermediate states along the FM paths from these representations. During training, we introduce an orthogonal projection method to adaptively determine the temporal position of these states, and apply a principled construction strategy based on a single-segment piecewise flow. The SFM inference starts from the intermediate state rather than pure noise, thereby focusing computation on the latter stages of the FM paths. We integrate SFM into multiple TTS models with a lightweight SFM head. Experiments demonstrate that SFM yields consistent gains in speech naturalness across both objective and subjective evaluations, and significantly accelerates inference when using adaptive-step ODE solvers. Demo and codes are available at https://ydqmkkx.github.io/SFMDemo/.

Yuki Saito, Shinnosuke Takamichi, Eiji Iimori et al.

We propose ChatGPT-EDSS, an empathetic dialogue speech synthesis (EDSS) method using ChatGPT for extracting dialogue context. ChatGPT is a chatbot that can deeply understand the content and purpose of an input prompt and appropriately respond to the user's request. We focus on ChatGPT's reading comprehension and introduce it to EDSS, a task of synthesizing speech that can empathize with the interlocutor's emotion. Our method first gives chat history to ChatGPT and asks it to generate three words representing the intention, emotion, and speaking style for each line in the chat. Then, it trains an EDSS model using the embeddings of ChatGPT-derived context words as the conditioning features. The experimental results demonstrate that our method performs comparably to ones using emotion labels or neural network-derived context embeddings learned from chat histories. The collected ChatGPT-derived context information is available at https://sarulab-speech.github.io/demo_ChatGPT_EDSS/.

Yuki Saito, Eiji Iimori, Shinnosuke Takamichi et al.

We present CALLS, a Japanese speech corpus that considers phone calls in a customer center as a new domain of empathetic spoken dialogue. The existing STUDIES corpus covers only empathetic dialogue between a teacher and student in a school. To extend the application range of empathetic dialogue speech synthesis (EDSS), we designed our corpus to include the same female speaker as the STUDIES teacher, acting as an operator in simulated phone calls. We describe a corpus construction methodology and analyze the recorded speech. We also conduct EDSS experiments using the CALLS and STUDIES corpora to investigate the effect of domain differences. The results show that mixing the two corpora during training causes biased improvements in the quality of synthetic speech due to the different degrees of expressiveness. Our project page of the corpus is http://sython.org/Corpus/STUDIES-2.

Kazuyuki Arikawa, Shoichi Koyama, Hiroshi Saruwatari

A spatial active noise control (ANC) method based on the individual kernel interpolation of primary and secondary sound fields is proposed. Spatial ANC is aimed at cancelling unwanted primary noise within a continuous region by using multiple secondary sources and microphones. A method based on the kernel interpolation of a sound field makes it possible to attenuate noise over the target region with flexible array geometry. Furthermore, by using the kernel function with directional weighting, prior information on primary noise source directions can be taken into consideration. However, whereas the sound field to be interpolated is a superposition of primary and secondary sound fields, the directional weight for the primary noise source was applied to the total sound field in previous work; therefore, the performance improvement was limited. We propose a method of individually interpolating the primary and secondary sound fields and formulate a normalized least-mean-square algorithm based on this interpolation method. Experimental results indicate that the proposed method outperforms the method based on total kernel interpolation.

Masaya Kawamura, Tomohiko Nakamura, Daichi Kitamura et al.

A differentiable digital signal processing (DDSP) autoencoder is a musical sound synthesizer that combines a deep neural network (DNN) and spectral modeling synthesis. It allows us to flexibly edit sounds by changing the fundamental frequency, timbre feature, and loudness (synthesis parameters) extracted from an input sound. However, it is designed for a monophonic harmonic sound and cannot handle mixtures of harmonic sounds. In this paper, we propose a model (DDSP mixture model) that represents a mixture as the sum of the outputs of multiple pretrained DDSP autoencoders. By fitting the output of the proposed model to the observed mixture, we can directly estimate the synthesis parameters of each source. Through synthesis parameter extraction experiments, we show that the proposed method has high and stable performance compared with a straightforward method that applies the DDSP autoencoder to the signals separated by an audio source separation method.

Keisuke Kimura, Shoichi Koyama, Natsuki Ueno et al.

A method of optimizing secondary source placement in sound field synthesis is proposed. Such an optimization method will be useful when the allowable placement region and available number of loudspeakers are limited. We formulate a mean-square-error-based cost function, incorporating the statistical properties of possible desired sound fields, for general linear-least-squares-based sound field synthesis methods, including pressure matching and (weighted) mode matching, whereas most of the current methods are applicable only to the pressure-matching method. An efficient greedy algorithm for minimizing the proposed cost function is also derived. Numerical experiments indicated that a high reproduction accuracy can be achieved by the placement optimized by the proposed method compared with the empirically used regular placement.

Ryosuke Horiuchi, Shoichi Koyama, Juliano G. C. Ribeiro et al.

A method to estimate an acoustic field from discrete microphone measurements is proposed. A kernel-interpolation-based method using the kernel function formulated for sound field interpolation has been used in various applications. The kernel function with directional weighting makes it possible to incorporate prior information on source directions to improve estimation accuracy. However, in prior studies, parameters for directional weighting have been empirically determined. We propose a method to optimize these parameters using observation values, which is particularly useful when prior information on source directions is uncertain. The proposed algorithm is based on discretization of the parameters and representation of the kernel function as a weighted sum of sub-kernels. Two types of regularization for the weights, $L_1$ and $L_2$, are investigated. Experimental results indicate that the proposed method achieves higher estimation accuracy than the method without kernel learning.

Takaaki Saeki, Shinnosuke Takamichi, Hiroshi Saruwatari

Incremental text-to-speech (TTS) synthesis generates utterances in small linguistic units for the sake of real-time and low-latency applications. We previously proposed an incremental TTS method that leverages a large pre-trained language model to take unobserved future context into account without waiting for the subsequent segment. Although this method achieves comparable speech quality to that of a method that waits for the future context, it entails a huge amount of processing for sampling from the language model at each time step. In this paper, we propose an incremental TTS method that directly predicts the unobserved future context with a lightweight model, instead of sampling words from the large-scale language model. We perform knowledge distillation from a GPT2-based context prediction network into a simple recurrent model by minimizing a teacher-student loss defined between the context embedding vectors of those models. Experimental results show that the proposed method requires about ten times less inference time to achieve comparable synthetic speech quality to that of our previous method, and it can perform incremental synthesis much faster than the average speaking speed of human English speakers, demonstrating the availability of our method to real-time applications.

Naoto Iijima, Shoichi Koyama, Hiroshi Saruwatari

A method of binaural rendering from microphone array signals of arbitrary geometry is proposed. To reproduce binaural signals from microphone array recordings at a remote location, a spherical microphone array is generally used for capturing a soundfield. However, owing to the lack of flexibility in the microphone arrangement, the single spherical array is sometimes impractical for estimating a large region of a soundfield. We propose a method based on harmonic analysis of infinite order, which allows the use of arbitrarily placed microphones. In the synthesis of the estimated soundfield, a spherical-wave-decomposition-based binaural rendering is also formulated to take into consideration the distance in measuring head-related transfer functions. We develop and evaluate a composite microphone array consisting of multiple small arrays. Experimental results including those of listening tests indicate that our proposed method is robust against change in listening position in the recording area.

Sota Misawa, Norihiro Takamune, Tomohiko Nakamura et al.

Rank-constrained spatial covariance matrix estimation (RCSCME) is a method for the situation that the directional target speech and the diffuse noise are mixed. In conventional RCSCME, independent low-rank matrix analysis (ILRMA) is used as the preprocessing method. We propose RCSCME using independent deeply learned matrix analysis (IDLMA), which is a supervised extension of ILRMA. In this method, IDLMA requires deep neural networks (DNNs) to separate the target speech and the noise. We use Denoiser, which is a single-channel speech enhancement DNN, in IDLMA to estimate not only the target speech but also the noise. We also propose noise self-supervised RCSCME, in which we estimate the noise-only time intervals using the output of Denoiser and design the prior distribution of the noise spatial covariance matrix for RCSCME. We confirm that the proposed methods outperform the conventional methods under several noise conditions.

Takuya Hasumi, Tomohiko Nakamura, Norihiro Takamune et al.

Independent deeply learned matrix analysis (IDLMA) is one of the state-of-the-art multichannel audio source separation methods using the source power estimation based on deep neural networks (DNNs). The DNN-based power estimation works well for sounds having timbres similar to the DNN training data. However, the sounds to which IDLMA is applied do not always have such timbres, and the timbral mismatch causes the performance degradation of IDLMA. To tackle this problem, we focus on a blind source separation counterpart of IDLMA, independent low-rank matrix analysis. It uses nonnegative matrix factorization (NMF) as the source model, which can capture source spectral components that only appear in the target mixture, using the low-rank structure of the source spectrogram as a clue. We thus extend the DNN-based source model to encompass the NMF-based source model on the basis of the product-of-expert concept, which we call the product of source models (PoSM). For the proposed PoSM-based IDLMA, we derive a computationally efficient parameter estimation algorithm based on an optimization principle called the majorization-minimization algorithm. Experimental evaluations show the effectiveness of the proposed method.

Yusaku Mizobuchi, Daichi Kitamura, Tomohiko Nakamura et al.

When we place microphones close to a sound source near other sources in audio recording, the obtained audio signal includes undesired sound from the other sources, which is often called cross-talk or bleeding sound. For many audio applications including onstage sound reinforcement and sound editing after a live performance, it is important to reduce the bleeding sound in each recorded signal. However, since microphones are spatially apart from each other in this situation, typical phase-aware blind source separation (BSS) methods cannot be used. We propose a phase-insensitive method for blind bleeding-sound reduction. This method is based on time-channel nonnegative matrix factorization, which is a BSS method using only amplitude spectrograms. With the proposed method, we introduce the gamma-distribution-based prior for leakage levels of bleeding sounds. Its optimization can be interpreted as maximum a posteriori estimation. The experimental results of music bleeding-sound reduction indicate that the proposed method is more effective for bleeding-sound reduction of music signals compared with other BSS methods.

Naoki Narisawa, Rintaro Ikeshita, Norihiro Takamune et al.

We address the determined audio source separation problem in the time-frequency domain. In independent deeply learned matrix analysis (IDLMA), it is assumed that the inter-frequency correlation of each source spectrum is zero, which is inappropriate for modeling nonstationary signals such as music signals. To account for the correlation between frequencies, independent positive semidefinite tensor analysis has been proposed. This unsupervised (blind) method, however, severely restrict the structure of frequency covariance matrices (FCMs) to reduce the number of model parameters. As an extension of these conventional approaches, we here propose a supervised method that models FCMs using deep neural networks (DNNs). It is difficult to directly infer FCMs using DNNs. Therefore, we also propose a new FCM model represented as a convex combination of a diagonal FCM and a rank-1 FCM. Our FCM model is flexible enough to not only consider inter-frequency correlation, but also capture the dynamics of time-varying FCMs of nonstationary signals. We infer the proposed FCMs using two DNNs: DNN for power spectrum estimation and DNN for time-domain signal estimation. An experimental result of separating music signals shows that the proposed method provides higher separation performance than IDLMA.

Takuya Hasumi, Tomohiko Nakamura, Norihiro Takamune et al.

Independent deeply learned matrix analysis (IDLMA) is one of the state-of-the-art supervised multichannel audio source separation methods. It blindly estimates the demixing filters on the basis of source independence, using the source model estimated by the deep neural network (DNN). However, since the ratios of the source to interferer signals vary widely among time-frequency (TF) slots, it is difficult to obtain reliable estimated power spectrograms of sources at all TF slots. In this paper, we propose an IDLMA extension, empirical Bayesian IDLMA (EB-IDLMA), by introducing a prior distribution of source power spectrograms and treating the source power spectrograms as latent random variables. This treatment allows us to implicitly consider the reliability of the estimated source power spectrograms for the estimation of demixing filters through the hyperparameters of the prior distribution estimated by the DNN. Experimental evaluations show the effectiveness of EB-IDLMA and the importance of introducing the reliability of the estimated source power spectrograms.

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.

Yuto Kondo, Yuki Kubo, Norihiro Takamune et al.

Rank-constrained spatial covariance matrix estimation (RCSCME) is a state-of-the-art blind speech extraction method applied to cases where one directional target speech and diffuse noise are mixed. In this paper, we proposed a new algorithmic extension of RCSCME. RCSCME complements a deficient one rank of the diffuse noise spatial covariance matrix, which cannot be estimated via preprocessing such as independent low-rank matrix analysis, and estimates the source model parameters simultaneously. In the conventional RCSCME, a direction of the deficient basis is fixed in advance and only the scale is estimated; however, the candidate of this deficient basis is not unique in general. In the proposed RCSCME model, the deficient basis itself can be accurately estimated as a vector variable by solving a vector optimization problem. Also, we derive new update rules based on the EM algorithm. We confirm that the proposed method outperforms conventional methods under several noise conditions.

Yota Ueda, Kazuki Fujii, Yuki Saito et al.

We propose a conditional generative adversarial network (GAN) incorporating humans' perceptual evaluations. A deep neural network (DNN)-based generator of a GAN can represent a real-data distribution accurately but can never represent a human-acceptable distribution, which are ranges of data in which humans accept the naturalness regardless of whether the data are real or not. A HumanGAN was proposed to model the human-acceptable distribution. A DNN-based generator is trained using a human-based discriminator, i.e., humans' perceptual evaluations, instead of the GAN's DNN-based discriminator. However, the HumanGAN cannot represent conditional distributions. This paper proposes the HumanACGAN, a theoretical extension of the HumanGAN, to deal with conditional human-acceptable distributions. Our HumanACGAN trains a DNN-based conditional generator by regarding humans as not only a discriminator but also an auxiliary classifier. The generator is trained by deceiving the human-based discriminator that scores the unconditioned naturalness and the human-based classifier that scores the class-conditioned perceptual acceptability. The training can be executed using the backpropagation algorithm involving humans' perceptual evaluations. Our experimental results in phoneme perception demonstrate that our HumanACGAN can successfully train this conditional generator.

Takaaki Saeki, Shinnosuke Takamichi, Hiroshi Saruwatari

This letter presents an incremental text-to-speech (TTS) method that performs synthesis in small linguistic units while maintaining the naturalness of output speech. Incremental TTS is generally subject to a trade-off between latency and synthetic speech quality. It is challenging to produce high-quality speech with a low-latency setup that does not make much use of an unobserved future sentence (hereafter, "lookahead"). To resolve this issue, we propose an incremental TTS method that uses a pseudo lookahead generated with a language model to take the future contextual information into account without increasing latency. Our method can be regarded as imitating a human's incremental reading and uses pretrained GPT2, which accounts for the large-scale linguistic knowledge, for the lookahead generation. Evaluation results show that our method 1) achieves higher speech quality than the method taking only observed information into account and 2) achieves a speech quality equivalent to waiting for the future context observation.

Shinnosuke Takamichi, Mamoru Komachi, Naoko Tanji et al.

In this paper, we construct a new Japanese speech corpus for speech-based summarization and simplification, "JSSS" (pronounced "j-triple-s"). Given the success of reading-style speech synthesis from short-form sentences, we aim to design more difficult tasks for delivering information to humans. Our corpus contains voices recorded for two tasks that have a role in providing information under constraints: duration-constrained text-to-speech summarization and speaking-style simplification. It also contains utterances of long-form sentences as an optional task. This paper describes how we designed the corpus, which is available on our project page.

Kentaro Mitsui, Tomoki Koriyama, Hiroshi Saruwatari

Multi-speaker speech synthesis is a technique for modeling multiple speakers' voices with a single model. Although many approaches using deep neural networks (DNNs) have been proposed, DNNs are prone to overfitting when the amount of training data is limited. We propose a framework for multi-speaker speech synthesis using deep Gaussian processes (DGPs); a DGP is a deep architecture of Bayesian kernel regressions and thus robust to overfitting. In this framework, speaker information is fed to duration/acoustic models using speaker codes. We also examine the use of deep Gaussian process latent variable models (DGPLVMs). In this approach, the representation of each speaker is learned simultaneously with other model parameters, and therefore the similarity or dissimilarity of speakers is considered efficiently. We experimentally evaluated two situations to investigate the effectiveness of the proposed methods. In one situation, the amount of data from each speaker is balanced (speaker-balanced), and in the other, the data from certain speakers are limited (speaker-imbalanced). Subjective and objective evaluation results showed that both the DGP and DGPLVM synthesize multi-speaker speech more effective than a DNN in the speaker-balanced situation. We also found that the DGPLVM outperforms the DGP significantly in the speaker-imbalanced situation.

Keigo Kamo, Yuki Kubo, Norihiro Takamune et al.

In this paper, we address a statistical model extension of multichannel nonnegative matrix factorization (MNMF) for blind source separation, and we propose a new parameter update algorithm used in the sub-Gaussian model. MNMF employs full-rank spatial covariance matrices and can simulate situations in which the reverberation is strong and the sources are not point sources. In conventional MNMF, spectrograms of observed signals are assumed to follow a multivariate Gaussian distribution. In this paper, first, to extend the MNMF model, we introduce the multivariate generalized Gaussian distribution as the multivariate sub-Gaussian distribution. Since the cost function of MNMF based on this multivariate sub-Gaussian model is difficult to minimize, we additionally introduce the joint-diagonalizability constraint in spatial covariance matrices to MNMF similarly to FastMNMF, and transform the cost function to the form to which we can apply the auxiliary functions to derive the valid parameter update rules. Finally, from blind source separation experiments, we show that the proposed method outperforms the conventional methods in source-separation accuracy.

Tomoki Koriyama, Hiroshi Saruwatari

This paper presents a deep Gaussian process (DGP) model with a recurrent architecture for speech sequence modeling. DGP is a Bayesian deep model that can be trained effectively with the consideration of model complexity and is a kernel regression model that can have high expressibility. In the previous studies, it was shown that the DGP-based speech synthesis outperformed neural network-based one, in which both models used a feed-forward architecture. To improve the naturalness of synthetic speech, in this paper, we show that DGP can be applied to utterance-level modeling using recurrent architecture models. We adopt a simple recurrent unit (SRU) for the proposed model to achieve a recurrent architecture, in which we can execute fast speech parameter generation by using the high parallelization nature of SRU. The objective and subjective evaluation results show that the proposed SRU-DGP-based speech synthesis outperforms not only feed-forward DGP but also automatically tuned SRU- and long short-term memory (LSTM)-based neural networks.

Tatsuki Kondo, Kanta Fukushige, Norihiro Takamune et al.

In this paper, we address a blind source separation (BSS) problem and propose a new extended framework of independent positive semidefinite tensor analysis (IPSDTA). IPSDTA is a state-of-the-art BSS method that enables us to take interfrequency correlations into account, but the generative model is limited within the multivariate Gaussian distribution and its parameter optimization algorithm does not guarantee stable convergence. To resolve these problems, first, we propose to extend the generative model to a parametric multivariate Student's t distribution that can deal with various types of signal. Secondly, we derive a new parameter optimization algorithm that guarantees the monotonic nonincrease in the cost function, providing stable convergence. Experimental results reveal that the cost function in the conventional IPSDTA does not display monotonically nonincreasing properties. On the other hand, the proposed method guarantees the monotonic nonincrease in the cost function and outperforms the conventional ILRMA and IPSDTA in the source-separation performance.

Takaaki Saeki, Yuki Saito, Shinnosuke Takamichi et al.

In this paper, we propose computationally efficient and high-quality methods for statistical voice conversion (VC) with direct waveform modification based on spectral differentials. The conventional method with a minimum-phase filter achieves high-quality conversion but requires heavy computation in filtering. This is because the minimum phase using a fixed lifter of the Hilbert transform often results in a long-tap filter. One of our methods is a data-driven method for lifter training. Since this method takes filter truncation into account in training, it can shorten the tap length of the filter while preserving conversion accuracy. Our other method is sub-band processing for extending the conventional method from narrow-band (16 kHz) to full-band (48 kHz) VC, which can convert a full-band waveform with higher converted-speech quality. Experimental results indicate that 1) the proposed lifter-training method for narrow-band VC can shorten the tap length to 1/16 without degrading the converted-speech quality and 2) the proposed sub-band-processing method for full-band VC can improve the converted-speech quality than the conventional method.

Keigo Kamo, Yuki Kubo, Norihiro Takamune et al.

In this paper, we address a convolutive blind source separation (BSS) problem and propose a new extended framework of FastMNMF by introducing prior information for joint diagonalization of the spatial covariance matrix model. Recently, FastMNMF has been proposed as a fast version of multichannel nonnegative matrix factorization under the assumption that the spatial covariance matrices of multiple sources can be jointly diagonalized. However, its source-separation performance was not improved and the physical meaning of the joint-diagonalization process was unclear. To resolve these problems, we first reveal a close relationship between the joint-diagonalization process and the demixing system used in independent low-rank matrix analysis (ILRMA). Next, motivated by this fact, we propose a new regularized FastMNMF supported by ILRMA and derive convergence-guaranteed parameter update rules. From BSS experiments, we show that the proposed method outperforms the conventional FastMNMF in source-separation accuracy with almost the same computation time.

Tomohiko Nakamura, Hiroshi Saruwatari

We propose a time-domain audio source separation method using down-sampling (DS) and up-sampling (US) layers based on a discrete wavelet transform (DWT). The proposed method is based on one of the state-of-the-art deep neural networks, Wave-U-Net, which successively down-samples and up-samples feature maps. We find that this architecture resembles that of multiresolution analysis, and reveal that the DS layers of Wave-U-Net cause aliasing and may discard information useful for the separation. Although the effects of these problems may be reduced by training, to achieve a more reliable source separation method, we should design DS layers capable of overcoming the problems. With this belief, focusing on the fact that the DWT has an anti-aliasing filter and the perfect reconstruction property, we design the proposed layers. Experiments on music source separation show the efficacy of the proposed method and the importance of simultaneously considering the anti-aliasing filters and the perfect reconstruction property.

Hiroki Tamaru, Shinnosuke Takamichi, Naoko Tanji et al.

Thanks to developments in machine learning techniques, it has become possible to synthesize high-quality singing voices of a single singer. An open multispeaker singing-voice corpus would further accelerate the research in singing-voice synthesis. However, conventional singing-voice corpora only consist of the singing voices of a single singer. We designed a Japanese multispeaker singing-voice corpus called "JVS-MuSiC" with the aim to analyze and synthesize a variety of voices. The corpus consists of 100 singers' recordings of the same song, Katatsumuri, which is a Japanese children's song. It also includes another song that is different for each singer. In this paper, we describe the design of the corpus and experimental analyses using JVS-MuSiC. We investigated the relationship between 1) the similarity of singing voices and perceptual oneness of unison singing voices and between 2) the similarity of singing voices and that of speech. The results suggest that 1) there is a positive and moderate correlation between singing-voice similarity and the oneness of unison and that 2) the correlation between singing-voice similarity and speech similarity is weak. This corpus is freely available online.

Kazuki Fujii, Yuki Saito, Shinnosuke Takamichi et al.

We propose the HumanGAN, a generative adversarial network (GAN) incorporating human perception as a discriminator. A basic GAN trains a generator to represent a real-data distribution by fooling the discriminator that distinguishes real and generated data. Therefore, the basic GAN cannot represent the outside of a real-data distribution. In the case of speech perception, humans can recognize not only human voices but also processed (i.e., a non-existent human) voices as human voice. Such a human-acceptable distribution is typically wider than a real-data one and cannot be modeled by the basic GAN. To model the human-acceptable distribution, we formulate a backpropagation-based generator training algorithm by regarding human perception as a black-boxed discriminator. The training efficiently iterates generator training by using a computer and discrimination by crowdsourcing. We evaluate our HumanGAN in speech naturalness modeling and demonstrate that it can represent a human-acceptable distribution that is wider than a real-data distribution.

Shinnosuke Takamichi, Kentaro Mitsui, Yuki Saito et al.

Thanks to improvements in machine learning techniques, including deep learning, speech synthesis is becoming a machine learning task. To accelerate speech synthesis research, we are developing Japanese voice corpora reasonably accessible from not only academic institutions but also commercial companies. In 2017, we released the JSUT corpus, which contains 10 hours of reading-style speech uttered by a single speaker, for end-to-end text-to-speech synthesis. For more general use in speech synthesis research, e.g., voice conversion and multi-speaker modeling, in this paper, we construct the JVS corpus, which contains voice data of 100 speakers in three styles (normal, whisper, and falsetto). The corpus contains 30 hours of voice data including 22 hours of parallel normal voices. This paper describes how we designed the corpus and summarizes the specifications. The corpus is available at our project page.

Yuki Kubo, Norihiro Takamune, Daichi Kitamura et al.

In this paper, we propose new accelerated update rules for rank-constrained spatial covariance model estimation, which efficiently extracts a directional target source in diffuse background noise.The naive updat e rule requires heavy computation such as matrix inversion or matrix multiplication. We resolve this problem by expanding matrix inversion to reduce computational complexity; in the parameter update step, we need neither matrix inversion nor multiplication. In an experiment, we show that the proposed accelerated update rule achieves 87 times faster calculation than the naive one.

Taiki Nakamura, Yuki Saito, Shinnosuke Takamichi et al.

This paper presents a new voice impersonation attack using voice conversion (VC). Enrolling personal voices for automatic speaker verification (ASV) offers natural and flexible biometric authentication systems. Basically, the ASV systems do not include the users' voice data. However, if the ASV system is unexpectedly exposed and hacked by a malicious attacker, there is a risk that the attacker will use VC techniques to reproduce the enrolled user's voices. We name this the ``verification-to-synthesis (V2S) attack'' and propose VC training with the ASV and pre-trained automatic speech recognition (ASR) models and without the targeted speaker's voice data. The VC model reproduces the targeted speaker's individuality by deceiving the ASV model and restores phonetic property of an input voice by matching phonetic posteriorgrams predicted by the ASR model. The experimental evaluation compares converted voices between the proposed method that does not use the targeted speaker's voice data and the standard VC that uses the data. The experimental results demonstrate that the proposed method performs comparably to the existing VC methods that trained using a very small amount of parallel voice data.

Yuki Saito, Shinnosuke Takamichi, Hiroshi Saruwatari

This paper proposes novel algorithms for speaker embedding using subjective inter-speaker similarity based on deep neural networks (DNNs). Although conventional DNN-based speaker embedding such as a $d$-vector can be applied to multi-speaker modeling in speech synthesis, it does not correlate with the subjective inter-speaker similarity and is not necessarily appropriate speaker representation for open speakers whose speech utterances are not included in the training data. We propose two training algorithms for DNN-based speaker embedding model using an inter-speaker similarity matrix obtained by large-scale subjective scoring. One is based on similarity vector embedding and trains the model to predict a vector of the similarity matrix as speaker representation. The other is based on similarity matrix embedding and trains the model to minimize the squared Frobenius norm between the similarity matrix and the Gram matrix of $d$-vectors, i.e., the inter-speaker similarity derived from the $d$-vectors. We crowdsourced the inter-speaker similarity scores of 153 Japanese female speakers, and the experimental results demonstrate that our algorithms learn speaker embedding that is highly correlated with the subjective similarity. We also apply the proposed speaker embedding to multi-speaker modeling in DNN-based speech synthesis and reveal that the proposed similarity vector embedding improves synthetic speech quality for open speakers whose speech utterances are unseen during the training.