張悦楷講古語音數據集

English

Dataset Description

• Homepage: 張悦楷講古語音數據集 The Zoeng Jyut Gaai Story-telling Speech Dataset
• License: CC0 1.0 Universal
• Language: Cantonese
• Total Duration: 188.25 hours
• Average Clip Duration: 5.826 seconds
• Median Clip Duration: 5.385 seconds
• Total number of characters: 2903094
• Average characters per clip: 24.90
• Median characters per clip: 23
• Average speech speed: 4.27 characters per second
• Voice Actor: 張悦楷

呢個係張悦楷講《三國演義》、《水滸傳》、《走進毛澤東的最後歲月》、《鹿鼎記》語音數據集。張悦楷係廣州最出名嘅講古佬 / 粵語説書藝人。佢從上世紀七十年代開始就喺廣東各個收音電台度講古，佢把聲係好多廣州人嘅共同回憶。本數據集收集嘅係佢最知名嘅四部作品。

數據集用途：

• TTS（語音合成）訓練集
• ASR（語音識別）訓練集或測試集
• 各種語言學、文學研究
• 直接聽嚟欣賞藝術！

TTS 效果演示：https://huggingface.co/spaces/laubonghaudoi/zoengjyutgaai_tts

説明

• 所有文本都根據 https://jyutping.org/blog/typo/ 同 https://jyutping.org/blog/particles/ 規範用字。
• 所有文本都使用全角標點，冇半角標點。
• 所有文本都用漢字轉寫，無阿拉伯數字無英文字母
• 所有音頻源都存放喺/source，為方便直接用作訓練數據，切分後嘅音頻都放喺 opus/
• 所有 opus 音頻皆為 48000 Hz 採樣率。
• 所有源字幕 SRT 文件都存放喺 srt/ 路經下，搭配 source/ 下嘅音源可以直接作為帶字幕嘅錄音直接欣賞。
• cut.py 係切分腳本，將對應嘅音源根據 srt 切分成短句並生成一個文本轉寫 csv。
• stats.py 係統計腳本，運行佢就會顯示成個數據集嘅各項統計數據。

引用本數據集

本數據集屬公共領域，遵循 CC0 許可聲明。即係話你可以無需授權免費任用本數據集，亦都唔需要註明出處。不過如果你用咗本數據集，我哋都希望你可以引用本頁面，作為對楷叔嘅懷念同致敬：

@misc{zoengjyutgaai2025,
    title={The Zoeng Jyut Gaai Story-telling Speech Dataset},
    author={Cantonese Computational Linguistics Infrastructure Development Workgroup (CanCLID)},
    howpublished = {\url{https://canclid.github.io/zoengjyutgaai/}},
    year={2025}
}

下載使用

要下載使用呢個數據集，可以喺 Python 入面直接跑：

from datasets import load_dataset

ds = load_dataset("CanCLID/zoengjyutgaai")

如果想單純將 opus/ 入面所有嘢下載落嚟，可以跑下面嘅 Python 代碼，注意要安裝 pip install --upgrade huggingface_hub 先：

from huggingface_hub import snapshot_download

# 如果淨係想下載啲字幕或者源音頻，就將 `opus/*` 改成 `srt/*` 或者 `source/*`
# If you only want to download subtitles or source audio, change `opus/*` to `srt/*` or `source/*`
snapshot_download(repo_id="CanCLID/zoengjyutgaai",allow_patterns="opus/*",local_dir="./",repo_type="dataset")

如果唔想用 python，你亦都可以用命令行叫 git 針對克隆個opus/或者其他路經，避免將成個 repo 都克隆落嚟浪費空間同下載時間：

mkdir zoengjyutgaai
cd zoengjyutgaai
git init
git lfs install

git remote add origin https://huggingface.co/datasets/CanCLID/zoengjyutgaai
git sparse-checkout init --cone

# 指定凈係下載個別路徑
git sparse-checkout set opus

# 開始下載
git pull origin main

數據集構建流程

本數據集嘅收集、構建過程係：

1. 從 YouTube 或者國內評書網站度下載錄音源文件，一般都係每集半個鐘長嘅 .webm 或者 .mp3。
2. 用加字幕工具幫呢啲錄音加字幕，得到對應嘅 .srt 文件。
3. 將啲源錄音用下面嘅命令儘可能無壓縮噉轉換成 .opus 格式。
4. 運行cut.py，將每一集 .opus 按照 .srt 入面嘅時間點切分成一句一個 .opus，然後對應嘅文本寫入本數據集嘅 xxx.csv。
5. 然後打開一個 IPython，逐句跑下面嘅命令，將啲數據推上 HuggingFace。

from datasets import load_dataset, DatasetDict
from huggingface_hub import login

sg = load_dataset('audiofolder', data_dir='./opus/saamgwokjinji')
sw = load_dataset('audiofolder', data_dir='./opus/seoiwuzyun')
mzd = load_dataset('audiofolder', data_dir='./opus/mouzaakdung')
ldg = load_dataset('audiofolder', data_dir='./opus/lukdinggei')
dataset = DatasetDict({
    "saamgwokjinji": sg["train"],
    "seoiwuzyun": sw["train"],
    "mouzaakdung": mzd["train"],
    "lukdinggei": ldg["train"],
})

# 檢查下讀入嘅數據有冇問題
dataset['mouzaakdung'][0]
# 準備好個 token 嚟登入
login()
# 推上 HuggingFace datasets
dataset.push_to_hub("CanCLID/zoengjyutgaai")

音頻格式轉換

首先要安裝 ffmpeg，然後運行：

# 將下載嘅音源由 webm 轉成 opus
ffmpeg -i webm/saamgwokjinji/001.webm -c:a copy source/saamgwokjinji/001.opus
# 或者轉 mp3
ffmpeg -i mp3/mouzaakdung/001.mp3 -c:a libopus -map_metadata -1 -b:a 48k -vbr on source/mouzaakdung/001.opus
# 將 opus 轉成無損 wav
ffmpeg -i source/saamgwokjinji/001.opus wav/saamgwokjinji/001.wav

如果想將所有 opus 文件全部轉換成 wav，可以直接運行to_wav.sh:

chmod +x to_wav.sh
./to_wav.sh

跟住就會生成一個 wav/ 路經，入面都係 opus/ 對應嘅音頻。注意 wav 格式非常掗埞，成個 opus/ 轉晒後會佔用至少 500GB 儲存空間，所以轉換之前記得確保有足夠空間。如果你想對音頻重採樣，亦都可以修改 to_wav.sh 入面嘅命令順便做重採樣。

The Zoeng Jyut Gaai Story-telling Speech Dataset

This is a speech dataset of Zoeng Jyut Gaai story-telling Romance of the Three Kingdoms, Water Margin and The Final Days of Mao Zedong. Zoeng Jyut Gaai is a famous actor, stand-up commedian and story-teller (講古佬) in 20th centry Canton. His voice remains in the memories of thousands of Cantonese people. This dataset is built from four of his most well-known story-telling pieces.

Use case of this dataset:

• TTS (Text-To-Speech) training set
• ASR (Automatic Speech Recognition) training or eval set
• Various linguistics / art analysis
• Just listen and enjoy the art piece!

TTS demo: https://huggingface.co/spaces/laubonghaudoi/zoengjyutgaai_tts

Introduction

• All transcriptions follow the prescribed orthography detailed in https://jyutping.org/blog/typo/ and https://jyutping.org/blog/particles/
• All transcriptions use full-width punctuations, no half-width punctuations is used.
• All transcriptions are in Chinese characters, no Arabic numbers or Latin letters.
• All source audio are stored in source/. For the convenice of training, segmented audios are stored in opus/.
• All opus audio are in 48000 Hz sampling rate.
• All source subtitle SRT files are stored in srt/. Use them with the webm files to enjoy subtitled storytelling pieces.
• cut.py is the script for cutting opus audios into senteneces based on the srt, and generates a csv file for transcriptions.
• stats.py is the script for getting stats of this dataset.

Citing this dataset

This dataset belongs to the public domain and follows the CC0 license agreement. This means you can use this dataset for free without attribution. However, if you use this dataset, we hope you can cite this page as a tribute to Gaai Suk:

@misc{zoengjyutgaai2025,
    title={The Zoeng Jyut Gaai Story-telling Speech Dataset},
    author={Cantonese Computational Linguistics Infrastructure Development Workgroup (CanCLID)},
    howpublished = {\url{https://canclid.github.io/zoengjyutgaai/}},
    year={2025}
}

Usage

To use this dataset, simply run in Python:

from datasets import load_dataset

ds = load_dataset("CanCLID/zoengjyutgaai")

If you only want to download a certain directory to save time and space from cloning the entire repo, run the Python codes below. Make sure you have pip install --upgrade huggingface_hub first：

from huggingface_hub import snapshot_download

# If you only want to download subtitles or source audio, change `opus/*` to `srt/*` or `source/*`
snapshot_download(repo_id="CanCLID/zoengjyutgaai",allow_patterns="opus/*",local_dir="./",repo_type="dataset")

If you don't want to run python codes and want to do this via command lines, you can selectively clone only a directory of the repo:

mkdir zoengjyutgaai
cd zoengjyutgaai
git init
git lfs install 

git remote add origin https://huggingface.co/datasets/CanCLID/zoengjyutgaai
git sparse-checkout init --cone

# Tell git which directory you want
git sparse-checkout set opus

# Pull the content
git pull origin main

Audio format conversion

Install ffmpeg first, then run：

# convert all webm into opus
ffmpeg -i webm/saamgwokjinji/001.webm -c:a copy source/saamgwokjinji/001.opus
# or into mp3
ffmpeg -i mp3/mouzaakdung/001.mp3 -c:a libopus -map_metadata -1 -b:a 48k -vbr on source/mouzaakdung/001.opus
# convert all opus into loseless wav
ffmpeg -i source/saamgwokjinji/001.opus wav/saamgwokjinji/001.wav

If you want to convert all opus to wav, run to_wav.sh:

chmod +x to_wav.sh
./to_wav.sh

It will generate a wav/ path which contains all audios converted from opus/. Be aware the wav format is very space-consuming. A full conversion will take up at least 500GB space so make sure you have enough storage. If you want to resample the audio, modify the line within to_wav.sh to resample the audio while doing the conversion.