README.md

---
license: cc
task_categories:
- text2text-generation
language:
- ja
tags:
- furigana
- education
pretty_name: 青空文庫振り仮名注釈付き音声コーパス(FLFL Train)
size_categories:
- 1M<n<10M
---

A more aggressively cleaned up version of [Calvin-Xu/Furigana-Aozora-Speech](https://huggingface.co/datasets/Calvin-Xu/Furigana-Aozora-Speech/blob/main/README.md), which consists of 2,536,041 out of the 3,361,443 entries generated from the raw data 青空文庫及びサピエの音声デイジーデータから作成した振り仮名注釈付き音声コーパスのデータセット https://github.com/ndl-lab/hurigana-speech-corpus-aozora. Training data of [Calvin-Xu/FLFL](https://huggingface.co/Calvin-Xu/FLFL).

The Whisper-generated transcriptions in the original dataset contains many errors. Additional sanity-checking is implemented to filter on reading of common kanji and eliminate wildly inaccurate entries (1,964,511 remaining).

Additionally as recognized by the original dataset authors, the consonant voicing (連濁) of many words, especially numbers / number + counters, is incorrect. Such words' readings were transformed in the training data (i.e., 十匹: じゅうひき -> じっぴき).

[Calvin-Xu/FLFL](https://huggingface.co/Calvin-Xu/FLFL/upload/main) is finetuned from [stockmark/gpt-neox-japanese-1.4b](https://huggingface.co/stockmark/gpt-neox-japanese-1.4b) and trained for slightly over one epoch on these data.

### FLFL Output Example

<ruby><ruby>国境<rt>くにざかい</rt></ruby>の<ruby>長<rt>なが</rt></ruby>いトンネルを<ruby>抜<rt>ぬ</rt></ruby>けると<ruby>雪国<rt>ゆきぐに</rt></ruby>であった<|endoftext|> 

<ruby>鰤<rt>ぶり</rt></ruby>の<ruby>照<rt>て</rt></ruby>り<ruby>焼<rt>や</rt></ruby>き、<ruby>八宝菜<rt>はっぽうさい</rt></ruby>、ハンバーグ。<|endoftext|> 

<ruby>主菜<rt>しゅさい</rt></ruby><ruby>関連<rt>かんれん</rt></ruby>は、<ruby>見事<rt>みごと</rt></ruby>なまでの<ruby>和洋<rt>わよう</rt></ruby><ruby>中<rt>ちゅう</rt></ruby><ruby>折衷<rt>せっちゅう</rt></ruby>。<|endoftext|> 

<ruby>別<rt>べつ</rt></ruby>の<ruby>者<rt>もの</rt></ruby>の<ruby>目<rt>め</rt></ruby>を<ruby>通<rt>つう</rt></ruby>じて<ruby>歴史<rt>れきし</rt></ruby>を<ruby>垣間見<rt>かいまみ</rt></ruby>られるとは、<ruby>想像<rt>そうぞう</rt></ruby>を<ruby>超<rt>こ</rt></ruby>える<ruby>体験<rt>たいけん</rt></ruby>に<ruby>違<rt>ちが</rt></ruby>いない!<|endoftext|> 

<ruby>止<rt>と</rt></ruby>めるなら、その<ruby>大本<rt>おおもと</rt></ruby>を<ruby>根絶<rt>ねだ</rt></ruby>やしにしないと<ruby>効果<rt>こうか</rt></ruby>がないわ<|endoftext|> 

<ruby>不人気<rt>ふにんき</rt></ruby><ruby>銘柄<rt>かぶ</rt></ruby>でこれ<ruby>以上<rt>いじょう</rt></ruby><ruby>価値<rt>かち</rt></ruby>が<ruby>下<rt>さ</rt></ruby>がりようないから、ほとんど<ruby>底値<rt>そこね</rt></ruby>だ<|endoftext|> 

<ruby>時間<rt>じかん</rt></ruby>の<ruby>澱<rt>おり</rt></ruby>の<ruby>中<rt>なか</rt></ruby>に<ruby>沈殿<rt>ちんたい</rt></ruby>していたようだ。<|endoftext|>