[Paper Review] JSUT corpus: free large-scale Japanese speech corpus for end-to-end speech synthesis

Ryosuke Sonobe, Shinnosuke Takamichi+1 more|arXiv (Cornell University)|Oct 28, 2017

Speech Recognition and Synthesis6 references88 citations

TL;DR

The paper introduces the JSUT corpus, a free 10-hour Japanese speech dataset designed to cover all daily-use kanji pronunciations for end-to-end speech synthesis, with nine sub-corpora and publicly available resources.

ABSTRACT

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.

Motivation & Objective

• Provide a free, large-scale Japanese speech corpus for end-to-end speech synthesis research.
• Ensure coverage of all pronunciations of daily-use Japanese characters and individual readings.
• Include diverse domains such as loanwords, paraphrase variants, and travel/precedent content.

Proposed method

• Design nine sub-corpora to cover main pronunciations and readings of daily-use kanji characters.
• Construct basic5000 by selecting 5000 sentences from Wikipedia and TANAKA corpus plus manually added sentences.
• Crowdsource or collect utterances including counter suffixes, loanwords, paraphrases, and domain-specific content.
• Incorporate para-speech from voice actress corpus and onomatopoeia sentences to enrich prosody.
• Record 10 hours of read speech from a native female speaker in an anechoic room at 48 kHz.
• Annotate and analyze linguistic statistics such as moras and words per utterance and measure F0 variation across days.

Experimental results

Research questions

• RQ1 Can a freely accessible Japanese speech corpus be constructed to cover all pronunciations of daily-use kanji characters for end-to-end TTS?
• RQ2 How do domain variations (loanwords, paraphrase, travel, precedent, etc.) affect data diversity and pronunciation coverage?
• RQ3 What are the linguistic and speech statistics of the constructed JSUT corpus and how do they vary over recording days?

Key findings

• The JSUT corpus comprises nine sub-corpora with varied linguistic content and 10 hours of speech data.
• The basic5000 sub-corpus targets all main pronunciations of daily-use kanji characters.
• The corpus includes loanwords, paraphrase variants, onomatopoeia, and domain-specific sentences to broaden coverage.
• Speech data were recorded in an anechoic room at 48 kHz from a native female speaker.
• The corpus provides UTF-8 text and 16-bit WAV speech data and is freely available online.
• Analysis shows a range of utterance lengths and an observable increase in mean log F0 over recording days.