JSUT corpus: free large-scale Japanese speech corpus for end-to-end speech synthesis

By Ryosuke Sonobe, Shinnosuke Takamichi, Hiroshi Saruwatari. 2017

185 citations recorded by the source corpus.

Thanks to improvements in machine learning techniques including deep learning, a free large-scale speech corpus that can be shared between academic institutions and commercial companies has an important role. However, such a corpus for Japanese speech synthesis does not exist. In this paper, we designed a novel Japanese speech corpus, named the "JSUT corpus," that is aimed at achieving end-to-end speech synthesis. The corpus consists of 10 hours of reading-style speech data and its transcription and covers all of the main pronunciations of daily-use Japanese characters. In this paper, we describe how we designed and analyzed the corpus. The corpus is freely available online.

Subfields: Artificial Intelligence

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

1. Xu Tan
2. Shinji Watanabe
3. Tao Qin
4. Tie-Yan Liu
5. Jinyu Li
6. Tomoki Hayashi
7. Lei He
8. Sheng Zhao

Citations in the corpus

• A Survey on Neural Speech Synthesis (2021, neutral): “1 Mandarin DiDiSpeech-1 [97] 572 4500 48 Mandarin DiDiSpeech-2 [97] 227 1500 48 Mandarin JSUT [308] 10 1 48 Japanese KazakhTTS [237] 93 2 44.”
• A Comparative Study on Transformer vs RNN in Speech Applications (2019, methodology): “Except for JSUT [25] and Fisher-CALLHOME Spanish, our data preparation scripts are based on Kaldi’s “s5x” recipe [7].”
• Unsupervised Learning For Sequence-to-sequence Text-to-speech For Low-resource Languages (2020, methodology): “For training VQ-VAE and pre-training Tacotron, we use open-source corpus in the following five languages: Korean [34], Japanese [35], Spanish, French, German 3 [36].”
• FoundationTTS: Text-to-Speech for ASR Customization with Generative Language Model (2023, background): “Although diversity can be achieved by sampling in inference time, these discrete models typically use very small amount and relatively clean dataset like popular JUST or LJ Speech [32, 33] , but large-scale and nosier datasets are not yet systemically explored especially in diverse speech output for”
• Espnet-TTS: Unified, Reproducible, and Integratable Open Source End-to-End Text-to-Speech Toolkit (2019, background): “[46] R. Sonobe, S. Takamichi, and H. Saruwatari, “JSUT corpus: Free large-scale Japanese speech corpus for end-to-end speech synthesis,” arXiv preprint arXiv:1711.00354, 2017.”
• CSS10: A Collection of Single Speaker Speech Datasets for 10 Languages (2019, background): “For the Japanese language, there is a public dataset called the JSUT dataset [19].”
• IndicSpeech: Text-to-Speech Corpus for Indian Languages (2020, background): “Some of these networks were also trained on datasets of comparable size for other languages like Korean (Park, 2019), Japanese (Sonobe et al., 2017) etc.”
• Phase Reconstruction from Amplitude Spectrograms Based on Von-Mises-Distribution Deep Neural Network (2018, methodology): “Experimental setup Evaluations were performed using the JSUT corpus [19], a free Japanese speech corpus uttered by a female speaker.”
• Neural iTTS: Toward Synthesizing Speech in Real-time with End-to-end Neural Text-to-Speech Framework (2019, methodology): “We used the JSUT 1.1 dataset that included 7,696 sentences (10 hours of audio sampled at 48-kHz, we downsampled to 22.05-kHz) spoken by a single native female speaker [25].”
• Open-source Multi-speaker Speech Corpora for Building Gujarati, Kannada, Malayalam, Marathi, Tamil and Telugu Speech Synthesis Systems (2020, background): “…approach, one would need to carefully design the recording script with the help of a linguist, recruit a voice talent, rent a professional studio and manage the recordings making sure the good quality is maintained throughout (Pitrelli et al., 2006; Ni et al., 2007; Sonobe et al., 2017).”