README.md

---
language:
- yue
license: cc0-1.0
size_categories:
- 10K<n<100K
task_categories:
- automatic-speech-recognition
- text-to-speech
- text-generation
- feature-extraction
- audio-to-audio
- audio-classification
- text-to-audio
pretty_name: c
configs:
- config_name: default
  data_files:
  - split: saamgwokjinji
    path: data/saamgwokjinji-*
  - split: seoiwuzyun
    path: data/seoiwuzyun-*
  - split: mouzaakdung
    path: data/mouzaakdung-*
tags:
- cantonese
- audio
- art
dataset_info:
  features:
  - name: audio
    dtype: audio
  - name: id
    dtype: string
  - name: episode_id
    dtype: int64
  - name: audio_duration
    dtype: float64
  - name: transcription
    dtype: string
  splits:
  - name: saamgwokjinji
    num_bytes: 2398591354.589
    num_examples: 39173
  - name: seoiwuzyun
    num_bytes: 1629539808.0
    num_examples: 24744
  - name: mouzaakdung
    num_bytes: 257168872.246
    num_examples: 4742
  download_size: 4304923024
  dataset_size: 4285300034.835
---

# 張悦楷講古語音數據集

[English](#the-zoeng-jyut-gaai-story-telling-speech-dataset)

## Dataset Description

- **Homepage:** [張悦楷講古語音數據集 The Zoeng Jyut Gaai Story-telling Speech Dataset](https://canclid.github.io/zoengjyutgaai/)
- **License:** [CC0 1.0 Universal](https://creativecommons.org/publicdomain/zero/1.0/)
- **Language:** Cantonese
- **Total Duration:** 112.54 hours
- **Average Clip Duration:** 5.901 seconds
- **Median Clip Duration:** 5.443 seconds
- **Total number of characters:** 1679097
- **Average characters per clip:** 24.36
- **Median characters per clip:** 23
- **Average speech speed:** 4.14 characters per second
- **Voice Actor:** [張悦楷](https://zh.wikipedia.org/wiki/%E5%BC%A0%E6%82%A6%E6%A5%B7)

呢個係張悦楷講《三國演義》、《水滸傳》、《走進毛澤東的最後歲月》語音數據集。[張悦楷](https://zh.wikipedia.org/wiki/%E5%BC%A0%E6%82%A6%E6%A5%B7)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三部作品。

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。

## 下載使用

要下載使用呢個數據集，可以喺 Python 入面直接跑：

```python
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：

```python
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：

```bash
mkdir zoengjyutgaai
cd zoengjyutgaai
git init

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`。
1. 用加字幕工具幫呢啲錄音加字幕，得到對應嘅 `.srt` 文件。
1. 將啲源錄音用下面嘅命令儘可能無壓縮噉轉換成 `.opus` 格式。
1. 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`。
1. 然後打開一個 IPython，逐句跑下面嘅命令，將啲數據推上 HuggingFace。

```python
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')
dataset = DatasetDict({
    "saamgwokjinji": sg["train"],
    "seoiwuzyun": sw["train"],
    "mouzaakdung": mzd["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](https://www.ffmpeg.org/download.html)，然後運行：

```bash
# 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](https://zh.wikipedia.org/wiki/%E5%BC%A0%E6%82%A6%E6%A5%B7) 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 three 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.

## Usage

To use this dataset, simply run in Python:

```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：

```python
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:

```bash
mkdir zoengjyutgaai
cd zoengjyutgaai
git init

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](https://www.ffmpeg.org/download.html) first, then run：

```bash
# 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.