Tomoki Hayashi

Brian Yan, Jiatong Shi, Yun Tang et al. · cmu, nvidia

ESPnet-ST-v2 is a revamp of the open-source ESPnet-ST toolkit necessitated by the broadening interests of the spoken language translation community. ESPnet-ST-v2 supports 1) offline speech-to-text translation (ST), 2) simultaneous speech-to-text translation (SST), and 3) offline speech-to-speech translation (S2ST) -- each task is supported with a wide variety of approaches, differentiating ESPnet-ST-v2 from other open source spoken language translation toolkits. This toolkit offers state-of-the-art architectures such as transducers, hybrid CTC/attention, multi-decoders with searchable intermediates, time-synchronous blockwise CTC/attention, Translatotron models, and direct discrete unit models. In this paper, we describe the overall design, example models for each task, and performance benchmarking behind ESPnet-ST-v2, which is publicly available at https://github.com/espnet/espnet.

Wen-Chin Huang, Shu-Wen Yang, Tomoki Hayashi et al.

We present a large-scale comparative study of self-supervised speech representation (S3R)-based voice conversion (VC). In the context of recognition-synthesis VC, S3Rs are attractive owing to their potential to replace expensive supervised representations such as phonetic posteriorgrams (PPGs), which are commonly adopted by state-of-the-art VC systems. Using S3PRL-VC, an open-source VC software we previously developed, we provide a series of in-depth objective and subjective analyses under three VC settings: intra-/cross-lingual any-to-one (A2O) and any-to-any (A2A) VC, using the voice conversion challenge 2020 (VCC2020) dataset. We investigated S3R-based VC in various aspects, including model type, multilinguality, and supervision. We also studied the effect of a post-discretization process with k-means clustering and showed how it improves in the A2A setting. Finally, the comparison with state-of-the-art VC systems demonstrates the competitiveness of S3R-based VC and also sheds light on the possible improving directions.

Massa Baali, Tomoki Hayashi, Hamdy Mubarak et al.

Several high-resource Text to Speech (TTS) systems currently produce natural, well-established human-like speech. In contrast, low-resource languages, including Arabic, have very limited TTS systems due to the lack of resources. We propose a fully unsupervised method for building TTS, including automatic data selection and pre-training/fine-tuning strategies for TTS training, using broadcast news as a case study. We show how careful selection of data, yet smaller amounts, can improve the efficiency of TTS system in generating more natural speech than a system trained on a bigger dataset. We adopt to propose different approaches for the: 1) data: we applied automatic annotations using DNSMOS, automatic vowelization, and automatic speech recognition (ASR) for fixing transcriptions' errors; 2) model: we used transfer learning from high-resource language in TTS model and fine-tuned it with one hour broadcast recording then we used this model to guide a FastSpeech2-based Conformer model for duration. Our objective evaluation shows 3.9% character error rate (CER), while the groundtruth has 1.3% CER. As for the subjective evaluation, where 1 is bad and 5 is excellent, our FastSpeech2-based Conformer model achieved a mean opinion score (MOS) of 4.4 for intelligibility and 4.2 for naturalness, where many annotators recognized the voice of the broadcaster, which proves the effectiveness of our proposed unsupervised method.

Tomoki Hayashi, Ryuichi Yamamoto, Takenori Yoshimura et al.

This paper describes ESPnet2-TTS, an end-to-end text-to-speech (E2E-TTS) toolkit. ESPnet2-TTS extends our earlier version, ESPnet-TTS, by adding many new features, including: on-the-fly flexible pre-processing, joint training with neural vocoders, and state-of-the-art TTS models with extensions like full-band E2E text-to-waveform modeling, which simplify the training pipeline and further enhance TTS performance. The unified design of our recipes enables users to quickly reproduce state-of-the-art E2E-TTS results. We also provide many pre-trained models in a unified Python interface for inference, offering a quick means for users to generate baseline samples and build demos. Experimental evaluations with English and Japanese corpora demonstrate that our provided models synthesize utterances comparable to ground-truth ones, achieving state-of-the-art TTS performance. The toolkit is available online at https://github.com/espnet/espnet.

Wen-Chin Huang, Shu-Wen Yang, Tomoki Hayashi et al.

This paper introduces S3PRL-VC, an open-source voice conversion (VC) framework based on the S3PRL toolkit. In the context of recognition-synthesis VC, self-supervised speech representation (S3R) is valuable in its potential to replace the expensive supervised representation adopted by state-of-the-art VC systems. Moreover, we claim that VC is a good probing task for S3R analysis. In this work, we provide a series of in-depth analyses by benchmarking on the two tasks in VCC2020, namely intra-/cross-lingual any-to-one (A2O) VC, as well as an any-to-any (A2A) setting. We also provide comparisons between not only different S3Rs but also top systems in VCC2020 with supervised representations. Systematic objective and subjective evaluation were conducted, and we show that S3R is comparable with VCC2020 top systems in the A2O setting in terms of similarity, and achieves state-of-the-art in S3R-based A2A VC. We believe the extensive analysis, as well as the toolkit itself, contribute to not only the S3R community but also the VC community. The codebase is now open-sourced.

Kazuhiro Kobayashi, Wen-Chin Huang, Yi-Chiao Wu et al.

In this paper, we present an open-source software for developing a nonparallel voice conversion (VC) system named crank. Although we have released an open-source VC software based on the Gaussian mixture model named sprocket in the last VC Challenge, it is not straightforward to apply any speech corpus because it is necessary to prepare parallel utterances of source and target speakers to model a statistical conversion function. To address this issue, in this study, we developed a new open-source VC software that enables users to model the conversion function by using only a nonparallel speech corpus. For implementing the VC software, we used a vector-quantized variational autoencoder (VQVAE). To rapidly examine the effectiveness of recent technologies developed in this research field, crank also supports several representative works for autoencoder-based VC methods such as the use of hierarchical architectures, cyclic architectures, generative adversarial networks, speaker adversarial training, and neural vocoders. Moreover, it is possible to automatically estimate objective measures such as mel-cepstrum distortion and pseudo mean opinion score based on MOSNet. In this paper, we describe representative functions developed in crank and make brief comparisons by objective evaluations.

Shinji Watanabe, Florian Boyer, Xuankai Chang et al.

This paper describes the recent development of ESPnet (https://github.com/espnet/espnet), an end-to-end speech processing toolkit. This project was initiated in December 2017 to mainly deal with end-to-end speech recognition experiments based on sequence-to-sequence modeling. The project has grown rapidly and now covers a wide range of speech processing applications. Now ESPnet also includes text to speech (TTS), voice conversation (VC), speech translation (ST), and speech enhancement (SE) with support for beamforming, speech separation, denoising, and dereverberation. All applications are trained in an end-to-end manner, thanks to the generic sequence to sequence modeling properties, and they can be further integrated and jointly optimized. Also, ESPnet provides reproducible all-in-one recipes for these applications with state-of-the-art performance in various benchmarks by incorporating transformer, advanced data augmentation, and conformer. This project aims to provide up-to-date speech processing experience to the community so that researchers in academia and various industry scales can develop their technologies collaboratively.

Chenda Li, Jing Shi, Wangyou Zhang et al.

We present ESPnet-SE, which is designed for the quick development of speech enhancement and speech separation systems in a single framework, along with the optional downstream speech recognition module. ESPnet-SE is a new project which integrates rich automatic speech recognition related models, resources and systems to support and validate the proposed front-end implementation (i.e. speech enhancement and separation).It is capable of processing both single-channel and multi-channel data, with various functionalities including dereverberation, denoising and source separation. We provide all-in-one recipes including data pre-processing, feature extraction, training and evaluation pipelines for a wide range of benchmark datasets. This paper describes the design of the toolkit, several important functionalities, especially the speech recognition integration, which differentiates ESPnet-SE from other open source toolkits, and experimental results with major benchmark datasets.

Pengcheng Guo, Florian Boyer, Xuankai Chang et al.

In this study, we present recent developments on ESPnet: End-to-End Speech Processing toolkit, which mainly involves a recently proposed architecture called Conformer, Convolution-augmented Transformer. This paper shows the results for a wide range of end-to-end speech processing applications, such as automatic speech recognition (ASR), speech translations (ST), speech separation (SS) and text-to-speech (TTS). Our experiments reveal various training tips and significant performance benefits obtained with the Conformer on different tasks. These results are competitive or even outperform the current state-of-art Transformer models. We are preparing to release all-in-one recipes using open source and publicly available corpora for all the above tasks with pre-trained models. Our aim for this work is to contribute to our research community by reducing the burden of preparing state-of-the-art research environments usually requiring high resources.

Wen-Chin Huang, Tomoki Hayashi, Shinji Watanabe et al.

This paper presents the sequence-to-sequence (seq2seq) baseline system for the voice conversion challenge (VCC) 2020. We consider a naive approach for voice conversion (VC), which is to first transcribe the input speech with an automatic speech recognition (ASR) model, followed using the transcriptions to generate the voice of the target with a text-to-speech (TTS) model. We revisit this method under a sequence-to-sequence (seq2seq) framework by utilizing ESPnet, an open-source end-to-end speech processing toolkit, and the many well-configured pretrained models provided by the community. Official evaluation results show that our system comes out top among the participating systems in terms of conversion similarity, demonstrating the promising ability of seq2seq models to convert speaker identity. The implementation is made open-source at: https://github.com/espnet/espnet/tree/master/egs/vcc20.

Hirofumi Inaguma, Shun Kiyono, Kevin Duh et al.

We present ESPnet-ST, which is designed for the quick development of speech-to-speech translation systems in a single framework. ESPnet-ST is a new project inside end-to-end speech processing toolkit, ESPnet, which integrates or newly implements automatic speech recognition, machine translation, and text-to-speech functions for speech translation. We provide all-in-one recipes including data pre-processing, feature extraction, training, and decoding pipelines for a wide range of benchmark datasets. Our reproducible results can match or even outperform the current state-of-the-art performances; these pre-trained models are downloadable. The toolkit is publicly available at https://github.com/espnet/espnet.

Tomoki Hayashi, Ryuichi Yamamoto, Katsuki Inoue et al.

This paper introduces a new end-to-end text-to-speech (E2E-TTS) toolkit named ESPnet-TTS, which is an extension of the open-source speech processing toolkit ESPnet. The toolkit supports state-of-the-art E2E-TTS models, including Tacotron~2, Transformer TTS, and FastSpeech, and also provides recipes inspired by the Kaldi automatic speech recognition (ASR) toolkit. The recipes are based on the design unified with the ESPnet ASR recipe, providing high reproducibility. The toolkit also provides pre-trained models and samples of all of the recipes so that users can use it as a baseline. Furthermore, the unified design enables the integration of ASR functions with TTS, e.g., ASR-based objective evaluation and semi-supervised learning with both ASR and TTS models. This paper describes the design of the toolkit and experimental evaluation in comparison with other toolkits. The experimental results show that our models can achieve state-of-the-art performance comparable to the other latest toolkits, resulting in a mean opinion score (MOS) of 4.25 on the LJSpeech dataset. The toolkit is publicly available at https://github.com/espnet/espnet.

Shigeki Karita, Nanxin Chen, Tomoki Hayashi et al.

Sequence-to-sequence models have been widely used in end-to-end speech processing, for example, automatic speech recognition (ASR), speech translation (ST), and text-to-speech (TTS). This paper focuses on an emergent sequence-to-sequence model called Transformer, which achieves state-of-the-art performance in neural machine translation and other natural language processing applications. We undertook intensive studies in which we experimentally compared and analyzed Transformer and conventional recurrent neural networks (RNN) in a total of 15 ASR, one multilingual ASR, one ST, and two TTS benchmarks. Our experiments revealed various training tips and significant performance benefits obtained with Transformer for each task including the surprising superiority of Transformer in 13/15 ASR benchmarks in comparison with RNN. We are preparing to release Kaldi-style reproducible recipes using open source and publicly available datasets for all the ASR, ST, and TTS tasks for the community to succeed our exciting outcomes.

Shinji Watanabe, Takaaki Hori, Shigeki Karita et al.

This paper introduces a new open source platform for end-to-end speech processing named ESPnet. ESPnet mainly focuses on end-to-end automatic speech recognition (ASR), and adopts widely-used dynamic neural network toolkits, Chainer and PyTorch, as a main deep learning engine. ESPnet also follows the Kaldi ASR toolkit style for data processing, feature extraction/format, and recipes to provide a complete setup for speech recognition and other speech processing experiments. This paper explains a major architecture of this software platform, several important functionalities, which differentiate ESPnet from other open source ASR toolkits, and experimental results with major ASR benchmarks.

Tatsuya Komatsu, Shinji Watanabe, Koichi Miyazaki et al.

This paper proposes acoustic event detection (AED) with classifier chains, a new classifier based on the probabilistic chain rule. The proposed AED with classifier chains consists of a gated recurrent unit and performs iterative binary detection of each event one by one. In each iteration, the event's activity is estimated and used to condition the next output based on the probabilistic chain rule to form classifier chains. Therefore, the proposed method can handle the interdependence among events upon classification, while the conventional AED methods with multiple binary classifiers with a linear layer and sigmoid function have placed an assumption of conditional independence. In the experiments with a real-recording dataset, the proposed method demonstrates its superior AED performance to a relative 14.80% improvement compared to a convolutional recurrent neural network baseline system with the multiple binary classifiers.

Jing Shi, Xuankai Chang, Tomoki Hayashi et al.

Deep learning based models have significantly improved the performance of speech separation with input mixtures like the cocktail party. Prominent methods (e.g., frequency-domain and time-domain speech separation) usually build regression models to predict the ground-truth speech from the mixture, using the masking-based design and the signal-level loss criterion (e.g., MSE or SI-SNR). This study demonstrates, for the first time, that the synthesis-based approach can also perform well on this problem, with great flexibility and strong potential. Specifically, we propose a novel speech separation/enhancement model based on the recognition of discrete symbols, and convert the paradigm of the speech separation/enhancement related tasks from regression to classification. By utilizing the synthesis model with the input of discrete symbols, after the prediction of discrete symbol sequence, each target speech could be re-synthesized. Evaluation results based on the WSJ0-2mix and VCTK-noisy corpora in various settings show that our proposed method can steadily synthesize the separated speech with high speech quality and without any interference, which is difficult to avoid in regression-based methods. In addition, with negligible loss of listening quality, the speaker conversion of enhanced/separated speech could be easily realized through our method.

Robin Karlsson, Tomoki Hayashi, Keisuke Fujii et al.

Recent self-supervised models have demonstrated equal or better performance than supervised methods, opening for AI systems to learn visual representations from practically unlimited data. However, these methods are typically classification-based and thus ineffective for learning high-resolution feature maps that preserve precise spatial information. This work introduces superpixels to improve self-supervised learning of dense semantically rich visual concept embeddings. Decomposing images into a small set of visually coherent regions reduces the computational complexity by $\mathcal{O}(1000)$ while preserving detail. We experimentally show that contrasting over regions improves the effectiveness of contrastive learning methods, extends their applicability to high-resolution images, improves overclustering performance, superpixels are better than grids, and regional masking improves performance. The expressiveness of our dense embeddings is demonstrated by improving the SOTA unsupervised semantic segmentation benchmark on Cityscapes, and for convolutional models on COCO.

Wen-Chin Huang, Tomoki Hayashi, Xinjian Li et al.

In voice conversion (VC), an approach showing promising results in the latest voice conversion challenge (VCC) 2020 is to first use an automatic speech recognition (ASR) model to transcribe the source speech into the underlying linguistic contents; these are then used as input by a text-to-speech (TTS) system to generate the converted speech. Such a paradigm, referred to as ASR+TTS, overlooks the modeling of prosody, which plays an important role in speech naturalness and conversion similarity. Although some researchers have considered transferring prosodic clues from the source speech, there arises a speaker mismatch during training and conversion. To address this issue, in this work, we propose to directly predict prosody from the linguistic representation in a target-speaker-dependent manner, referred to as target text prediction (TTP). We evaluate both methods on the VCC2020 benchmark and consider different linguistic representations. The results demonstrate the effectiveness of TTP in both objective and subjective evaluations.

Ibuki Kuroyanagi, Tomoki Hayashi, Kazuya Takeda et al.

An anomalous sound detection system to detect unknown anomalous sounds usually needs to be built using only normal sound data. Moreover, it is desirable to improve the system by effectively using a small amount of anomalous sound data, which will be accumulated through the system's operation. As one of the methods to meet these requirements, we focus on a binary classification model that is developed by using not only normal data but also outlier data in the other domains as pseudo-anomalous sound data, which can be easily updated by using anomalous data. In this paper, we implement a new loss function based on metric learning to learn the distance relationship from each class centroid in feature space for the binary classification model. The proposed multi-task learning of the binary classification and the metric learning makes it possible to build the feature space where the within-class variance is minimized and the between-class variance is maximized while keeping normal and anomalous classes linearly separable. We also investigate the effectiveness of additionally using anomalous sound data for further improving the binary classification model. Our results showed that multi-task learning using binary classification and metric learning to consider the distance from each class centroid in the feature space is effective, and performance can be significantly improved by using even a small amount of anomalous data during training.

Tomoki Hayashi, Wen-Chin Huang, Kazuhiro Kobayashi et al.

This paper proposes a novel voice conversion (VC) method based on non-autoregressive sequence-to-sequence (NAR-S2S) models. Inspired by the great success of NAR-S2S models such as FastSpeech in text-to-speech (TTS), we extend the FastSpeech2 model for the VC problem. We introduce the convolution-augmented Transformer (Conformer) instead of the Transformer, making it possible to capture both local and global context information from the input sequence. Furthermore, we extend variance predictors to variance converters to explicitly convert the source speaker's prosody components such as pitch and energy into the target speaker. The experimental evaluation with the Japanese speaker dataset, which consists of male and female speakers of 1,000 utterances, demonstrates that the proposed model enables us to perform more stable, faster, and better conversion than autoregressive S2S (AR-S2S) models such as Tacotron2 and Transformer.

Wen-Chin Huang, Yi-Chiao Wu, Tomoki Hayashi et al.

We present a novel approach to any-to-one (A2O) voice conversion (VC) in a sequence-to-sequence (seq2seq) framework. A2O VC aims to convert any speaker, including those unseen during training, to a fixed target speaker. We utilize vq-wav2vec (VQW2V), a discretized self-supervised speech representation that was learned from massive unlabeled data, which is assumed to be speaker-independent and well corresponds to underlying linguistic contents. Given a training dataset of the target speaker, we extract VQW2V and acoustic features to estimate a seq2seq mapping function from the former to the latter. With the help of a pretraining method and a newly designed postprocessing technique, our model can be generalized to only 5 min of data, even outperforming the same model trained with parallel data.

Wen-Chin Huang, Tomoki Hayashi, Yi-Chiao Wu et al.

Sequence-to-sequence (seq2seq) voice conversion (VC) models are attractive owing to their ability to convert prosody. Nonetheless, without sufficient data, seq2seq VC models can suffer from unstable training and mispronunciation problems in the converted speech, thus far from practical. To tackle these shortcomings, we propose to transfer knowledge from other speech processing tasks where large-scale corpora are easily available, typically text-to-speech (TTS) and automatic speech recognition (ASR). We argue that VC models initialized with such pretrained ASR or TTS model parameters can generate effective hidden representations for high-fidelity, highly intelligible converted speech. We apply such techniques to recurrent neural network (RNN)-based and Transformer based models, and through systematical experiments, we demonstrate the effectiveness of the pretraining scheme and the superiority of Transformer based models over RNN-based models in terms of intelligibility, naturalness, and similarity.

Yi-Chiao Wu, Tomoki Hayashi, Takuma Okamoto et al.

In this paper, we propose a quasi-periodic parallel WaveGAN (QPPWG) waveform generative model, which applies a quasi-periodic (QP) structure to a parallel WaveGAN (PWG) model using pitch-dependent dilated convolution networks (PDCNNs). PWG is a small-footprint GAN-based raw waveform generative model, whose generation time is much faster than real time because of its compact model and non-autoregressive (non-AR) and non-causal mechanisms. Although PWG achieves high-fidelity speech generation, the generic and simple network architecture lacks pitch controllability for an unseen auxiliary fundamental frequency ($F_{0}$) feature such as a scaled $F_{0}$. To improve the pitch controllability and speech modeling capability, we apply a QP structure with PDCNNs to PWG, which introduces pitch information to the network by dynamically changing the network architecture corresponding to the auxiliary $F_{0}$ feature. Both objective and subjective experimental results show that QPPWG outperforms PWG when the auxiliary $F_{0}$ feature is scaled. Moreover, analyses of the intermediate outputs of QPPWG also show better tractability and interpretability of QPPWG, which respectively models spectral and excitation-like signals using the cascaded fixed and adaptive blocks of the QP structure.

Yi-Chiao Wu, Tomoki Hayashi, Patrick Lumban Tobing et al.

In this paper, a pitch-adaptive waveform generative model named Quasi-Periodic WaveNet (QPNet) is proposed to improve the limited pitch controllability of vanilla WaveNet (WN) using pitch-dependent dilated convolution neural networks (PDCNNs). Specifically, as a probabilistic autoregressive generation model with stacked dilated convolution layers, WN achieves high-fidelity audio waveform generation. However, the pure-data-driven nature and the lack of prior knowledge of audio signals degrade the pitch controllability of WN. For instance, it is difficult for WN to precisely generate the periodic components of audio signals when the given auxiliary fundamental frequency ($F_{0}$) features are outside the $F_{0}$ range observed in the training data. To address this problem, QPNet with two novel designs is proposed. First, the PDCNN component is applied to dynamically change the network architecture of WN according to the given auxiliary $F_{0}$ features. Second, a cascaded network structure is utilized to simultaneously model the long- and short-term dependencies of quasi-periodic signals such as speech. The performances of single-tone sinusoid and speech generations are evaluated. The experimental results show the effectiveness of the PDCNNs for unseen auxiliary $F_{0}$ features and the effectiveness of the cascaded structure for speech generation.

Yi-Chiao Wu, Tomoki Hayashi, Takuma Okamoto et al.

In this paper, we propose a parallel WaveGAN (PWG)-like neural vocoder with a quasi-periodic (QP) architecture to improve the pitch controllability of PWG. PWG is a compact non-autoregressive (non-AR) speech generation model, whose generative speed is much faster than real time. While utilizing PWG as a vocoder to generate speech on the basis of acoustic features such as spectral and prosodic features, PWG generates high-fidelity speech. However, when the input acoustic features include unseen pitches, the pitch accuracy of PWG-generated speech degrades because of the fixed and generic network of PWG without prior knowledge of speech periodicity. The proposed QPPWG adopts a pitch-dependent dilated convolution network (PDCNN) module, which introduces the pitch information into PWG via the dynamically changed network architecture, to improve the pitch controllability and speech modeling capability of vanilla PWG. Both objective and subjective evaluation results show the higher pitch accuracy and comparable speech quality of QPPWG-generated speech when the QPPWG model size is only 70 % of that of vanilla PWG.

Tomoki Hayashi, Shinji Watanabe

This paper proposes a new end-to-end text-to-speech (E2E-TTS) model based on neural machine translation (NMT). The proposed model consists of two components; a non-autoregressive vector quantized variational autoencoder (VQ-VAE) model and an autoregressive Transformer-NMT model. The VQ-VAE model learns a mapping function from a speech waveform into a sequence of discrete symbols, and then the Transformer-NMT model is trained to estimate this discrete symbol sequence from a given input text. Since the VQ-VAE model can learn such a mapping in a fully-data-driven manner, we do not need to consider hyperparameters of the feature extraction required in the conventional E2E-TTS models. Thanks to the use of discrete symbols, we can use various techniques developed in NMT and automatic speech recognition (ASR) such as beam search, subword units, and fusions with a language model. Furthermore, we can avoid an over smoothing problem of predicted features, which is one of the common issues in TTS. The experimental evaluation with the JSUT corpus shows that the proposed method outperforms the conventional Transformer-TTS model with a non-autoregressive neural vocoder in naturalness, achieving the performance comparable to the reconstruction of the VQ-VAE model.

Yi-Chiao Wu, Patrick Lumban Tobing, Kazuhiro Kobayashi et al.

In this paper, we integrate a simple non-parallel voice conversion (VC) system with a WaveNet (WN) vocoder and a proposed collapsed speech suppression technique. The effectiveness of WN as a vocoder for generating high-fidelity speech waveforms on the basis of acoustic features has been confirmed in recent works. However, when combining the WN vocoder with a VC system, the distorted acoustic features, acoustic and temporal mismatches, and exposure bias usually lead to significant speech quality degradation, making WN generate some very noisy speech segments called collapsed speech. To tackle the problem, we take conventional-vocoder-generated speech as the reference speech to derive a linear predictive coding distribution constraint (LPCDC) to avoid the collapsed speech problem. Furthermore, to mitigate the negative effects introduced by the LPCDC, we propose a collapsed speech segment detector (CSSD) to ensure that the LPCDC is only applied to the problematic segments to limit the loss of quality to short periods. Objective and subjective evaluations are conducted, and the experimental results confirm the effectiveness of the proposed method, which further improves the speech quality of our previous non-parallel VC system submitted to Voice Conversion Challenge 2018.

Takenori Yoshimura, Tomoki Hayashi, Kazuya Takeda et al.

This paper integrates a voice activity detection (VAD) function with end-to-end automatic speech recognition toward an online speech interface and transcribing very long audio recordings. We focus on connectionist temporal classification (CTC) and its extension of CTC/attention architectures. As opposed to an attention-based architecture, input-synchronous label prediction can be performed based on a greedy search with the CTC (pre-)softmax output. This prediction includes consecutive long blank labels, which can be regarded as a non-speech region. We use the labels as a cue for detecting speech segments with simple thresholding. The threshold value is directly related to the length of a non-speech region, which is more intuitive and easier to control than conventional VAD hyperparameters. Experimental results on unsegmented data show that the proposed method outperformed the baseline methods using the conventional energy-based and neural-network-based VAD methods and achieved an RTF less than 0.2. The proposed method is publicly available.

Wen-Chin Huang, Tomoki Hayashi, Yi-Chiao Wu et al.

We introduce a novel sequence-to-sequence (seq2seq) voice conversion (VC) model based on the Transformer architecture with text-to-speech (TTS) pretraining. Seq2seq VC models are attractive owing to their ability to convert prosody. While seq2seq models based on recurrent neural networks (RNNs) and convolutional neural networks (CNNs) have been successfully applied to VC, the use of the Transformer network, which has shown promising results in various speech processing tasks, has not yet been investigated. Nonetheless, their data-hungry property and the mispronunciation of converted speech make seq2seq models far from practical. To this end, we propose a simple yet effective pretraining technique to transfer knowledge from learned TTS models, which benefit from large-scale, easily accessible TTS corpora. VC models initialized with such pretrained model parameters are able to generate effective hidden representations for high-fidelity, highly intelligible converted speech. Experimental results show that such a pretraining scheme can facilitate data-efficient training and outperform an RNN-based seq2seq VC model in terms of intelligibility, naturalness, and similarity.

Patrick Lumban Tobing, Yi-Chiao Wu, Tomoki Hayashi et al.

In this paper, we present a novel technique for a non-parallel voice conversion (VC) with the use of cyclic variational autoencoder (CycleVAE)-based spectral modeling. In a variational autoencoder(VAE) framework, a latent space, usually with a Gaussian prior, is used to encode a set of input features. In a VAE-based VC, the encoded latent features are fed into a decoder, along with speaker-coding features, to generate estimated spectra with either the original speaker identity (reconstructed) or another speaker identity (converted). Due to the non-parallel modeling condition, the converted spectra can not be directly optimized, which heavily degrades the performance of a VAE-based VC. In this work, to overcome this problem, we propose to use CycleVAE-based spectral model that indirectly optimizes the conversion flow by recycling the converted features back into the system to obtain corresponding cyclic reconstructed spectra that can be directly optimized. The cyclic flow can be continued by using the cyclic reconstructed features as input for the next cycle. The experimental results demonstrate the effectiveness of the proposed CycleVAE-based VC, which yields higher accuracy of converted spectra, generates latent features with higher correlation degree, and significantly improves the quality and conversion accuracy of the converted speech.

Yi-Chiao Wu, Patrick Lumban Tobing, Tomoki Hayashi et al.

In this paper, we investigate the effectiveness of a quasi-periodic WaveNet (QPNet) vocoder combined with a statistical spectral conversion technique for a voice conversion task. The WaveNet (WN) vocoder has been applied as the waveform generation module in many different voice conversion frameworks and achieves significant improvement over conventional vocoders. However, because of the fixed dilated convolution and generic network architecture, the WN vocoder lacks robustness against unseen input features and often requires a huge network size to achieve acceptable speech quality. Such limitations usually lead to performance degradation in the voice conversion task. To overcome this problem, the QPNet vocoder is applied, which includes a pitch-dependent dilated convolution component to enhance the pitch controllability and attain a more compact network than the WN vocoder. In the proposed method, input spectral features are first converted using a framewise deep neural network, and then the QPNet vocoder generates converted speech conditioned on the linearly converted prosodic and transformed spectral features. The experimental results confirm that the QPNet vocoder achieves significantly better performance than the same-size WN vocoder while maintaining comparable speech quality to the double-size WN vocoder. Index Terms: WaveNet, vocoder, voice conversion, pitch-dependent dilated convolution, pitch controllability

Yi-Chiao Wu, Tomoki Hayashi, Patrick Lumban Tobing et al.

In this paper, we propose a quasi-periodic neural network (QPNet) vocoder with a novel network architecture named pitch-dependent dilated convolution (PDCNN) to improve the pitch controllability of WaveNet (WN) vocoder. The effectiveness of the WN vocoder to generate high-fidelity speech samples from given acoustic features has been proved recently. However, because of the fixed dilated convolution and generic network architecture, the WN vocoder hardly generates speech with given F0 values which are outside the range observed in training data. Consequently, the WN vocoder lacks the pitch controllability which is one of the essential capabilities of conventional vocoders. To address this limitation, we propose the PDCNN component which has the time-variant adaptive dilation size related to the given F0 values and a cascade network structure of the QPNet vocoder to generate quasi-periodic signals such as speech. Both objective and subjective tests are conducted, and the experimental results demonstrate the better pitch controllability of the QPNet vocoder compared to the same and double sized WN vocoders while attaining comparable speech qualities. Index Terms: WaveNet, vocoder, quasi-periodic signal, pitch-dependent dilated convolution, pitch controllability

Wen-Chin Huang, Yi-Chiao Wu, Chen-Chou Lo et al.

In this work, we investigate the effectiveness of two techniques for improving variational autoencoder (VAE) based voice conversion (VC). First, we reconsider the relationship between vocoder features extracted using the high quality vocoders adopted in conventional VC systems, and hypothesize that the spectral features are in fact F0 dependent. Such hypothesis implies that during the conversion phase, the latent codes and the converted features in VAE based VC are in fact source F0 dependent. To this end, we propose to utilize the F0 as an additional input of the decoder. The model can learn to disentangle the latent code from the F0 and thus generates converted F0 dependent converted features. Second, to better capture temporal dependencies of the spectral features and the F0 pattern, we replace the frame wise conversion structure in the original VAE based VC framework with a fully convolutional network structure. Our experiments demonstrate that the degree of disentanglement as well as the naturalness of the converted speech are indeed improved.

Wen-Chin Huang, Yi-Chiao Wu, Hsin-Te Hwang et al.

This paper presents a refinement framework of WaveNet vocoders for variational autoencoder (VAE) based voice conversion (VC), which reduces the quality distortion caused by the mismatch between the training data and testing data. Conventional WaveNet vocoders are trained with natural acoustic features but conditioned on the converted features in the conversion stage for VC, and such a mismatch often causes significant quality and similarity degradation. In this work, we take advantage of the particular structure of VAEs to refine WaveNet vocoders with the self-reconstructed features generated by VAE, which are of similar characteristics with the converted features while having the same temporal structure with the target natural features. We analyze these features and show that the self-reconstructed features are similar to the converted features. Objective and subjective experimental results demonstrate the effectiveness of our proposed framework.

Takaaki Hori, Ramon Astudillo, Tomoki Hayashi et al.

This paper presents a method to train end-to-end automatic speech recognition (ASR) models using unpaired data. Although the end-to-end approach can eliminate the need for expert knowledge such as pronunciation dictionaries to build ASR systems, it still requires a large amount of paired data, i.e., speech utterances and their transcriptions. Cycle-consistency losses have been recently proposed as a way to mitigate the problem of limited paired data. These approaches compose a reverse operation with a given transformation, e.g., text-to-speech (TTS) with ASR, to build a loss that only requires unsupervised data, speech in this example. Applying cycle consistency to ASR models is not trivial since fundamental information, such as speaker traits, are lost in the intermediate text bottleneck. To solve this problem, this work presents a loss that is based on the speech encoder state sequence instead of the raw speech signal. This is achieved by training a Text-To-Encoder model and defining a loss based on the encoder reconstruction error. Experimental results on the LibriSpeech corpus show that the proposed cycle-consistency training reduced the word error rate by 14.7% from an initial model trained with 100-hour paired data, using an additional 360 hours of audio data without transcriptions. We also investigate the use of text-only data mainly for language modeling to further improve the performance in the unpaired data training scenario.

Tomoki Hayashi, Shinji Watanabe, Yu Zhang et al.

In this paper we propose a novel data augmentation method for attention-based end-to-end automatic speech recognition (E2E-ASR), utilizing a large amount of text which is not paired with speech signals. Inspired by the back-translation technique proposed in the field of machine translation, we build a neural text-to-encoder model which predicts a sequence of hidden states extracted by a pre-trained E2E-ASR encoder from a sequence of characters. By using hidden states as a target instead of acoustic features, it is possible to achieve faster attention learning and reduce computational cost, thanks to sub-sampling in E2E-ASR encoder, also the use of the hidden states can avoid to model speaker dependencies unlike acoustic features. After training, the text-to-encoder model generates the hidden states from a large amount of unpaired text, then E2E-ASR decoder is retrained using the generated hidden states as additional training data. Experimental evaluation using LibriSpeech dataset demonstrates that our proposed method achieves improvement of ASR performance and reduces the number of unknown words without the need for paired data.

Yi-Chiao Wu, Kazuhiro Kobayashi, Tomoki Hayashi et al.

In this paper, we propose a technique to alleviate the quality degradation caused by collapsed speech segments sometimes generated by the WaveNet vocoder. The effectiveness of the WaveNet vocoder for generating natural speech from acoustic features has been proved in recent works. However, it sometimes generates very noisy speech with collapsed speech segments when only a limited amount of training data is available or significant acoustic mismatches exist between the training and testing data. Such a limitation on the corpus and limited ability of the model can easily occur in some speech generation applications, such as voice conversion and speech enhancement. To address this problem, we propose a technique to automatically detect collapsed speech segments. Moreover, to refine the detected segments, we also propose a waveform generation technique for WaveNet using a linear predictive coding constraint. Verification and subjective tests are conducted to investigate the effectiveness of the proposed techniques. The verification results indicate that the detection technique can detect most collapsed segments. The subjective evaluations of voice conversion demonstrate that the generation technique significantly improves the speech quality while maintaining the same speaker similarity.

Tomoki Hayashi, Shinji Watanabe, Tomoki Toda et al.

This paper presents a new network architecture called multi-head decoder for end-to-end speech recognition as an extension of a multi-head attention model. In the multi-head attention model, multiple attentions are calculated, and then, they are integrated into a single attention. On the other hand, instead of the integration in the attention level, our proposed method uses multiple decoders for each attention and integrates their outputs to generate a final output. Furthermore, in order to make each head to capture the different modalities, different attention functions are used for each head, leading to the improvement of the recognition performance with an ensemble effect. To evaluate the effectiveness of our proposed method, we conduct an experimental evaluation using Corpus of Spontaneous Japanese. Experimental results demonstrate that our proposed method outperforms the conventional methods such as location-based and multi-head attention models, and that it can capture different speech/linguistic contexts within the attention-based encoder-decoder framework.