This paper integrates a classic mel-cepstral synthesis filter into a modern neural speech synthesis system towards end-to-end controllable speech synthesis. Since the mel-cepstral synthesis filter is explicitly embedded in neural waveform models in the proposed system, both voice characteristics and the pitch of synthesized speech are highly controlled via a frequency warping parameter and fundamental frequency, respectively. We implement the mel-cepstral synthesis filter as a differentiable and GPU-friendly module to enable the acoustic and waveform models in the proposed system to be simultaneously optimized in an end-to-end manner. Experiments show that the proposed system improves speech quality from a baseline system maintaining controllability. The core PyTorch modules used in the experiments are publicly available on GitHub1.

We propose PeriodNet, a non-autoregressive (non-AR) waveform generation model with a new model structure for modeling periodic and aperiodic components in speech waveforms. The non-AR waveform generation models can generate speech waveforms parallelly and can be used as a speech vocoder by conditioning an acoustic feature. Since a speech waveform contains periodic and aperiodic components, both components should be appropriately modeled to generate a high-quality speech waveform. However, it is difficult to decompose the components from a natural speech waveform in advance. To address this issue, we propose a parallel model and a series model structure separating periodic and aperiodic components. The features of our proposed models are that explicit periodic and aperiodic signals are taken as input, and external periodic/aperiodic decomposition is not needed in training. Experiments using a singing voice corpus show that our proposed structure improves the naturalness of the generated waveform. We also show that the speech waveforms with a pitch outside of the training data range can be generated with more naturalness.

This paper presents Sinsy, a deep neural network (DNN)-based singing voice synthesis (SVS) system. In recent years, DNNs have been utilized in statistical parametric SVS systems, and DNN-based SVS systems have demonstrated better performance than conventional hidden Markov model-based ones. SVS systems are required to synthesize a singing voice with pitch and timing that strictly follow a given musical score. Additionally, singing expressions that are not described on the musical score, such as vibrato and timing fluctuations, should be reproduced. The proposed system is composed of four modules: a time-lag model, a duration model, an acoustic model, and a vocoder, and singing voices can be synthesized taking these characteristics of singing voices into account. To better model a singing voice, the proposed system incorporates improved approaches to modeling pitch and vibrato and better training criteria into the acoustic model. In addition, we incorporated PeriodNet, a non-autoregressive neural vocoder with robustness for the pitch, into our systems to generate a high-fidelity singing voice waveform. Moreover, we propose automatic pitch correction techniques for DNN-based SVS to synthesize singing voices with correct pitch even if the training data has out-of-tune phrases. Experimental results show our system can synthesize a singing voice with better timing, more natural vibrato, and correct pitch, and it can achieve better mean opinion scores in subjective evaluation tests.

This paper proposes a general framework of semi-supervised learning based on hierarchical generative models and adapts it to a Japanese end-to-end text-to-speech (TTS) system. In English TTS, several end-to-end systems have recently achieved sound quality close to that of natural human speech. However, in non-alphabetic languages such as Japanese, it is difficult to realize true text-input end-to-end TTS due to character diversity and pitch accents. To address this problem, we propose end-to-end TTS based on semi-supervised learning that makes the most of existing data consisting of any combination of text, phoneme, and waveform as training data. To demonstrate the effectiveness of the proposed system, listening tests were conducted for pronunciation and naturalness. Our results show that the proposed system improves both pronunciation and naturalness.

The present paper describes singing voice synthesis based on convolutional neural networks (CNNs). Singing voice synthesis systems based on deep neural networks (DNNs) are currently being proposed and are improving the naturalness of synthesized singing voices. As singing voices represent a rich form of expression, a powerful technique to model them accurately is required. In the proposed technique, long-term dependencies of singing voices are modeled by CNNs. An acoustic feature sequence is generated for each segment that consists of long-term frames, and a natural trajectory is obtained without the parameter generation algorithm. Furthermore, a computational complexity reduction technique, which drives the DNNs in different time units depending on type of musical score features, is proposed. Experimental results show that the proposed method can synthesize natural sounding singing voices much faster than the conventional method.

This paper proposes a hierarchical generative model with a multi-grained latent variable to synthesize expressive speech. In recent years, fine-grained latent variables are introduced into the text-to-speech synthesis that enable the fine control of the prosody and speaking styles of synthesized speech. However, the naturalness of speech degrades when these latent variables are obtained by sampling from the standard Gaussian prior. To solve this problem, we propose a novel framework for modeling the fine-grained latent variables, considering the dependence on an input text, a hierarchical linguistic structure, and a temporal structure of latent variables. This framework consists of a multi-grained variational autoencoder, a conditional prior, and a multi-level auto-regressive latent converter to obtain the different time-resolution latent variables and sample the finer-level latent variables from the coarser-level ones by taking into account the input text. Experimental results indicate an appropriate method of sampling fine-grained latent variables without the reference signal at the synthesis stage. Our proposed framework also provides the controllability of speaking style in an entire utterance.

This paper proposes a generative adversarial training method for deep neural network (DNN)-based singing voice synthesis. The DNN-based approach has been used in statistical parametric singing voice synthesis and improved the naturalness of the synthesized singing voice [1]. Recently, generative adversarial networks (GANs) [2] have attracted significant attention in various machine learning research areas including speech synthesis [3]. GANs have achieved great success in modeling the distributions of complex data, and they have the potential to alleviate over-smoothing problem on the generated speech parameters in speech synthesis. In this paper, we propose a DNN-based singing voice synthesis system incorporating the GAN. Experimental results show that the proposed method outperforms the conventional method in the naturalness of the synthesized singing voice.

This paper proposes a WaveNet-based delay-free adaptive differential pulse code modulation (ADPCM) speech coding system. The WaveNet generative model, which is a state-of-the-art model for neural-network-based speech waveform synthesis, is used as the adaptive predictor in ADPCM. To further improve speech quality, mel-cepstrum-based noise shaping and postfiltering were integrated with the proposed ADPCM system. Both objective and subjective evaluation results indicate that the proposed ADPCM system outperformed not only the conventional ADPCM system based on ITU-T Recommendation G.726 but also the ADPCM system based on adaptive mel-cepstral analysis.

This paper proposes a voice conversion technique based on WaveNet to directly generate target audio waveforms from acoustic features of a source speaker. In voice conversion based on statistical models, the relation between acoustic features, such as spectral parameters, extracted from source and target audio waveforms is generally modeled using statistical models, such as Gaussian mixture models and neural networks. Although modeling the relation between acoustic features is reasonable and efficient, these models are not optimized for predicting target audio waveforms because the vocoder parameters are used as intermediate representations. To overcome this problem, we developed a voice conversion method to model the relation between target audio waveforms and acoustic features extracted from source audio waveforms using WaveNet, which is a generative model for audio waveforms. The proposed model can directly generate converted audio waveforms without vocoders. Experimental results indicate that the proposed method can generate a more naturally sounding converted speech than that using a conventional DNN method.

This paper presents a WaveNet-based zero-delay lossless speech coding technique for high-quality communications. The WaveNet generative model, which is a state-of-the-art model for neural-network-based speech waveform synthesis, is used in both the encoder and decoder. In the encoder, discrete speech signals are losslessly compressed using sample-by-sample entropy coding. The decoder fully reconstructs the original speech signals from the compressed signals without algorithmic delay. Experimental results show that the proposed coding technique can transmit speech audio waveforms with 50% their original bit rate and the WaveNet-based speech coder remains effective for unknown speakers.

This paper presents a mel-cepstrum-based quantization noise shaping method for improving the quality of synthetic speech generated by neural-network-based speech waveform synthesis systems. Since mel-cepstral coefficients closely match the characteristics of human auditory perception, the proposed method effectively masks the white noise introduced by the quantization typically used in neural-network-based speech waveform synthesis systems. The paper also describes a computationally efficient implementation of the proposed method using the structure of the mel-log spectrum approximation filter. Experiments using the WaveNet generative model, which is a state-of-the-art model for neural-network-based speech waveform synthesis, showed that speech quality is significantly improved by the proposed method.

This paper proposes a novel method to build multiple decision trees as a structure of factor analyzed hidden Markov model for speech synthesis. In the proposed method, the multiple decision trees grow simultaneously rather than sequentially to take into account the relationship between the trees. However, the simultaneous growing is computationally infeasible due to an exponential increase in the number of tree structures to be evaluated. To solve the problem, we further propose two computational complexity reduction algorithms that achieve a significant reduction in the computational time. Experimental results show that the proposed method outperforms the conventional one based on a single decision tree.

This paper proposes a new training method of deep neural networks (DNNs) for statistical parametric speech synthesis. DNNs are recently used as acoustic models that represent mapping functions from linguistic features to acoustic features in statistical parametric speech synthesis. There are problems to be solved in conventional DNN-based speech synthesis: 1) the inconsistency between the training and synthesis criteria; and 2) the over-smoothing of the generated parameter trajectories. In this paper, we introduce the parameter trajectory generation process considering the global variance (GV) into the training of DNNs. A unified framework which consistently uses the same criterion in both training and synthesis can be obtained and the model parameters are optimized for parameter generation considering the GV in the proposed method. Experimental results show that the proposed method outperforms the conventional method in the naturalness of synthesized speech.

This paper investigates how to use neural networks in statistical parametric speech synthesis. Recently, deep neural networks (DNNs) have been used for statistical parametric speech synthesis. However, the specific way how DNNs should be used in statistical parametric speech synthesis has not been studied thoroughly. A generation process of statistical parametric speech synthesis based on generative models can be divided into several components, and those components can be represented by DNNs. In this paper, the effect of DNNs for each component is investigated by comparing DNNs with generative models. Experimental results show that the use of a DNN as acoustic models is effective and the parameter generation combined with a DNN improves the naturalness of synthesized speech.

In this paper, we propose a voice interaction system using 3D-CG virtual agents for stand-alone smartphones. Because the proposed system can handle speech recognition and speech synthesis on a stand-alone smartphone differently from the existing mobile voice interaction systems, this system enables us to talk naturally without encountering delays caused by network communications. Moreover, proposed system can be fully customized by dialogue scripts, Java-based plugins, and Android APIs. Therefore, developers can make original voice interaction systems for smartphones easily based on proposed system. We have made a subset of the proposed system available as open-source software. We expect that this system will contribute to studies of human-agent interaction using smartphones.

The present paper describes Japanese and English singing voice synthesis systems based on hidden Markov models (HMMs). In this approach, the spectrum, excitation, and vibrato of the singing voice are simultaneously modeled by context-dependent HMMs, and waveforms are generated by the HMMs themselves. Japanese singing voice synthesis systems have already been developed and used to create variable musical contents. To extend this system to English, language independent contexts are designed. Furthermore, methods for matching musical notes and pronunciation of English lyrics are presented and evaluated in subjective experiments. Then, Japanese and English singing voice synthesis systems are compared.

A statistical parametric approach to singing voice synthesis based on hidden Markov models (HMMs) has been growing in popularity over the last few years. The spectrum, excitation, vibrato, and duration of the singing voice in this approach are simultaneously modeled with context-dependent HMMs and waveforms are generated from the HMMs themselves. Since HMM-based singing voice synthesis systems are “corpus-based,” the HMMs corresponding to contextual factors that rarely appear in the training data cannot be well-trained. However, it may be difficult to prepare a large enough quantity of singing voice data sung by one singer. Furthermore, the pitch included in each song is imbalanced, and there is the vocal range of the singer. In this paper, we propose “singer adaptive training” which can solve the data sparse-ness problem. Experimental results demonstrated that the proposed technique improved the quality of the synthesized singing voices.

This paper describes development of an open-source toolkit which makes it possible to explore a vast variety of aspects in speech interactions at spoken dialog systems and speech interfaces. The toolkit tightly incorporates recent speech recognition and synthesis technologies with a 3-D CG rendering module that can manipulates expressive embodied agent characters. The software design and its interfaces are carefully designed to be fully open toolkit. Ongoing demonstration experiments to public indicates that it is promoting related researches and developments of voice interaction systems in various scenes.

This paper gives a general overview of hidden Markov model (HMM)-based speech synthesis, which has recently been demonstrated to be very effective in synthesizing speech. The main advantage of this approach is its flexibility in changing speaker identities, emotions, and speaking styles. This paper also discusses the relation between the HMM-based approach and the more conventional unit-selection approach that has dominated over the last decades. Finally, advanced techniques for future developments are described.

A statistical parametric approach to singing voice synthesis based on hidden Markov Models (HMMs) has been growing in popularity over the last few years. The spectrum, excitation, vibrato, and duration of singing voices in this approach are simultaneously modeled with context-dependent HMMs and waveforms are generated from the HMMs themselves. HMM-based singing voice synthesis systems are heavily based on the training data in performance because these systems are “corpus-based.” Therefore, HMMs corresponding to contextual factors that hardly ever appear in the training data cannot be well-trained. Pitch should especially be correctly covered since generated F0trajectories have a great impact on the subjective quality of synthesized singing voices. We applied the method of “speaker adaptive training” (SAT) to “pitch adaptive training,” which is discussed in this paper. This technique made it possible to normalize pitch based on musical notes in the training process. The experimental results demonstrated that the proposed technique could alleviate the data sparseness problem.

This paper proposes a technique for constructing independent parameter tying structures of mean and variance in HMM based speech synthesis. Conventionally, mean and variance parameters are assumed to have the same tying structure. However, it has been reported that a clustering technique of mean vectors while tying all variance matrices improves the quality of synthesized speech. This indicates that mean and variance parameters should have different optimal tying structures. In the proposed technique, the decision trees for mean and variance parameters are simultaneously grown by taking into account the dependency on mean and variance parameters. Experimental results show that the proposed technique outperforms the conventional one.

Three speech rate control methods for HMM-based speech synthesis were compared by large-scale subjective evaluations. The methods are 1) synthesizing speech sounds based on HMMs trained from corpora at a target speech rate, 2) stretching or shrinking utterance durations proportionally in waveform generation, and 3) determining state durations based on ML criterion under a restriction of utterance duration. The results indicated that the proportional shrinking had significant advantages for fast rate, whereas HMMs trained from slow speech sounds had a slight advantage for slow rate. We also found an advantage of proportionally shrunk speech from a synthesizer trained from slow speech corpora.

In the EMIME project, we are developing a mobile device that performs personalized speech-to-speech translation such that a user's spoken input in one language is used to produce spoken output in another language, while continuing to sound like the user's voice. We integrate two techniques, unsupervised adaptation for HMM-based TTS using a word-based large-vocabulary continuous speech recognizer and cross-lingual speaker adaptation for HMM-based TTS, into a single architecture. Thus, an unsupervised cross-lingual speaker adaptation system can be developed. Listening tests show very promising results, demonstrating that adapted voices sound similar to the target speaker and that differences between supervised and unsupervised cross-lingual speaker adaptation are small.

In conventional speech synthesis, large amounts of phonetically balanced speech data recorded in highly controlled recording studio environments are typically required to build a voice. Although using such data is a straightforward solution for high quality synthesis, the number of voices available will always be limited, because recording costs are high. On the other hand, our recent experiments with HMM-based speech synthesis systems have demonstrated that speaker-adaptive HMM-based speech synthesis (which uses an “average voice model” plus model adaptation) is robust to non-ideal speech data that are recorded under various conditions and with varying microphones, that are not perfectly clean, and/or that lack phonetic balance. This enables us to consider building high-quality voices on “non-TTS” corpora such as ASR corpora. Since ASR corpora generally include a large number of speakers, this leads to the possibility of producing an enormous number of voices automatically. In this paper, we demonstrate the thousands of voices for HMM-based speech synthesis that we have made from several popular ASR corpora such as the Wall Street Journal (WSJ0, WSJ1, and WSJCAM0), Resource Management, Globalphone, and SPEECON databases. We also present the results of associated analysis based on perceptual evaluation, and discuss remaining issues.

This paper proposes a technique of reducing footprint of HMMbased speech synthesis systems by tying all covariance matrices. HMM-based speech synthesis systems usually consume smaller footprint than unit-selection synthesis systems because statistics rather than speech waveforms are stored. However, further reduction is essential to put them on embedded devices which have very small memory. According to the empirical knowledge that covariance matrices have smaller impact for the quality of synthesized speech than mean vectors, here we propose a clustering technique of mean vectors while tying all covariance matrices. Subjective listening test results show that the proposed technique can shrink the footprint of an HMM-based speech synthesis system while retaining the quality of synthesized speech. Index Terms: HMM, speech synthesis, decision tree, contextclustering, MDL criterion, embedded device

Our recent experiments with HMM-based speech synthesis systems have demonstrated that speaker-adaptive HMM-based speech synthesis (which uses an 'average voice model' plus model adaptation) is robust to non-ideal speech data that are recorded under various conditions and with varying microphones, that are not perfectly clean, and/or that lack of phonetic balance. This enables us consider building high-quality voices on 'non-TTS' corpora such as ASR corpora. Since ASR corpora generally include a large number of speakers, this leads to the possibility of producing an enormous number of voices automatically. In this paper we show thousands of voices for HMM-based speech synthesis that we have made from several popular ASR corpora such as the Wall Street Journal databases (WSJ0/WSJ1/WSJCAM0), Resource Management, Globalphone and Speecon. We report some perceptual evaluation results and outline the outstanding issues.

In hidden Markov models (HMMs), state duration probabilities decrease exponentially with time. It would be inappropriate representation of temporal structure of speech. One of the solutions for this problem is integrating state duration probability distributions explicitly into the HMM. This form is known as a hidden semi-Markov model (HSMM) [1]. Although a number of attempts to use explicit duration models in speech recognition systems have been proposed, they are not consistent because various approximations were used in both training and decoding. In the present paper, a fully consistent speech recognition system based on the HSMM framework is proposed. In a speaker-dependent continuous speech recognition experiment, HSMM-based speech recognition system achieved about 5.9% relative error reduction over the corresponding HMM-based one.