% Manual for xjyutping.  Compile with: xelatex xjyutping-doc.tex (twice)
\documentclass[11pt]{article}
\usepackage[a4paper,margin=25mm]{geometry}
\usepackage{xeCJK}
\setCJKmainfont{Songti TC}
\usepackage{xcolor}
\usepackage{booktabs}
\usepackage{fancyvrb}
\usepackage[hidelinks]{hyperref}
\usepackage{xjyutping}

\setlength\emergencystretch{3em}
\newcommand\pkg[1]{\textsf{#1}}
\newcommand\cs[1]{\texttt{\textbackslash#1}}
\newcommand\meta[1]{\ensuremath{\langle}\textit{#1}\ensuremath{\rangle}}
\newcommand\marg[1]{\texttt{\{}\meta{#1}\texttt{\}}}
\newcommand\oarg[1]{\texttt{[}\meta{#1}\texttt{]}}
% A code sample followed by its output.
\newenvironment{example}
  {\par\smallskip\VerbatimEnvironment\begin{Verbatim}[frame=single,fontsize=\small]}
  {\end{Verbatim}\par\noindent\ignorespacesafterend}

\title{\pkg{xjyutping}: Jyutping above Traditional Chinese characters}
\author{Version 1.5.0}
\date{29 September 2026}

\begin{document}
\maketitle

\begin{abstract}
\noindent
This package adds Cantonese Jyutping (粵拼), the romanisation of Cantonese
devised by the Linguistic Society of Hong Kong (LSHK), above Traditional
Chinese characters. It is aimed at anyone preparing Cantonese material in
\LaTeX{}, such as teaching notes, lyrics or poetry, where the pronunciation of
each character should be shown above it. The reading of each character is
chosen from the surrounding words, so 行 is read \emph{hong4} in 銀行 but
\emph{haang4} in 行路. Every character sits in a cell of the same width, which
is wide enough for its Jyutping, so the text stays evenly spaced and no
Jyutping touches its neighbours or the line above. The package works with
Xe\LaTeX{} through \pkg{xeCJK} and with Lua\LaTeX{} through \pkg{LuaTeX-ja},
including the \pkg{ctex} package and classes under either engine.
\end{abstract}

\section{Getting started}

\noindent
To annotate a document,
\begin{enumerate}
\item Load \pkg{ctex}, \pkg{xeCJK} or \pkg{luatexja-fontspec} and choose a
  Traditional Chinese font
\item Load \pkg{xjyutping}
\item Put the text inside a \texttt{jyutpingscope} environment
\item Compile with \texttt{xelatex} or \texttt{lualatex}
\end{enumerate}

\noindent
A minimal document is,
\begin{example}
\documentclass{article}
\usepackage[fontset=none]{ctex}    % for xelatex or lualatex
\setCJKmainfont{Songti TC}
\usepackage{xjyutping}
\begin{document}
\begin{jyutpingscope}
我哋去銀行，行路返屋企。校長喺學校長大。
\end{jyutpingscope}
\end{document}
\end{example}

\begin{jyutpingscope}
我哋去銀行，行路返屋企。校長喺學校長大。
\end{jyutpingscope}

\noindent
The \pkg{ctex} package sets up Chinese typesetting for either engine. Instead,
\pkg{xeCJK} can be loaded under Xe\LaTeX{}, or \pkg{luatexja-fontspec} (with
\cs{setmainjfont}) under Lua\LaTeX{}. If none of them is loaded,
\pkg{xjyutping} loads \pkg{xeCJK} or \pkg{LuaTeX-ja} itself. Any Traditional
Chinese font can be used, and the Jyutping is set in the main Latin font of
the document unless another font is chosen (Section~\ref{sec:options}).

\section{Commands}

The commands are mainly divided into two groups, the commands that annotate
text (Section~\ref{sec:annotate}) and the commands that choose a reading
(Section~\ref{sec:choose}).

\subsection{Annotating text}\label{sec:annotate}

\begin{description}
\item[\cs{begin}\{jyutpingscope\}\oarg{options} \dots\ \cs{end}\{jyutpingscope\}]
  annotates every Chinese character in the block. The block ends the
  paragraph, and its line spacing is raised where needed so that the Jyutping
  of one line never comes close to the line above.
\item[\cs{xjyutping*}\oarg{options}\marg{text}] does the same for a piece of
  running text inside an ordinary paragraph.
\end{description}

\noindent
For example,
\begin{example}
香港人講\xjyutping*{廣東話}，寫\xjyutping*{繁體字}。
\end{example}
香港人講\xjyutping*{廣東話}，寫\xjyutping*{繁體字}。

\subsection{Choosing a reading yourself}\label{sec:choose}

Cantonese pronunciation changes with meaning and with colloquial tone change,
and no word list is perfect. Hence there are three ways to set a reading by
hand,

\begin{description}
\item[\cs{xjyutping}\oarg{options}\marg{characters}\marg{readings}] gives the
  characters an explicit reading at that point only. It works inside a scope
  and on its own, and it also marks a word boundary.
\item[\cs{setjyutping}\marg{character}\marg{reading}] changes the reading a
  character takes when it is not part of a known word.
\item[\cs{setjyutping}\marg{word}\marg{readings}] adds a word or overrides the
  reading of a known one. Words always take priority over the defaults of
  single characters.
\end{description}

\noindent
Readings are Jyutping syllables separated by spaces, with one syllable per
Chinese character. Punctuation, Latin letters and commands among the
characters take no syllable, so \verb|\xjyutping{A行}{hong4}| reads 行 as
\emph{hong4}. Since \cs{setjyutping} is global and takes effect from where it
appears, it can go in the preamble or in the middle of a scope.

\noindent
In the following example, 重話 means `also said', so 重 is read \emph{zung6}
rather than with its default \emph{cung5} (`heavy'), while 家行 is a name
read \emph{gaa1 hang4},
\begin{example}
\setjyutping{重話}{zung6 waa6}
\begin{jyutpingscope}
佢重話\xjyutping{家行}{gaa1 hang4}聽日嚟。
\end{jyutpingscope}
\end{example}
\setjyutping{重話}{zung6 waa6}
\begin{jyutpingscope}
佢重話\xjyutping{家行}{gaa1 hang4}聽日嚟。
\end{jyutpingscope}

\subsection{Formatting inside words}

Braces and formatting commands do not split a word, so a character can be
highlighted without losing its reading in context. For example,
\begin{example}
\begin{jyutpingscope}
佢喺銀\textbf{行}返工，係校\textcolor{red}{長}嘅朋友。
\end{jyutpingscope}
\end{example}
\begin{jyutpingscope}
佢喺銀\textbf{行}返工，係校\textcolor{red}{長}嘅朋友。
\end{jyutpingscope}

\subsection{Switching off}

The \cs{disablejyutping} command stops annotation until \cs{enablejyutping}
or the end of the current group or environment. Use it for passages inside a
scope that should stay plain.

\section{How readings are chosen}

The reading of each character is chosen in five steps,
\begin{enumerate}
\item First, the text is split into runs of Chinese characters. Punctuation,
  Latin text and most commands end a run, while spaces and line breaks
  between two characters do not, and neither do braces or formatting
  commands (\cs{textbf}, \cs{emph}, \cs{color}, \cs{textcolor}, size commands
  \dots). The text of a \cs{footnote} is a run of its own, and the text
  around the footnote mark carries on.
\item Then each run is divided into words from a list of about 104\,000
  Cantonese words. The division with the fewest words is chosen, then the
  one with the fewest single characters, then the one of the more common
  words, as counted in the word-frequency list of rime-cantonese and in the
  transcripts of WenetSpeech-Yue (步行街 is 步行 + 街, not 步 + 行街). Only a
  full tie goes to the longer final word.
\item A word takes its reading from the list, and the words set with
  \cs{setjyutping} are checked first and win a tie of step 2.
\item A character that is not part of a word takes the reading set with
  \cs{setjyutping} if there is one. Otherwise, some characters take their
  reading from the one or two characters after them, a few are read
  differently at the end of a run, and the rest take their default reading.
  For example, 呢 is the demonstrative \emph{ni1} before a classifier
  (呢張床) but the particle \emph{ne1} in 你呢？ and 佢呢就走, while 咁 is
  \emph{gam2} ``like this'' at the end of a run (就係咁) but \emph{gam3}
  ``so'' before an adjective (咁大).
\item Lastly, a reading given with
  \cs{xjyutping}\marg{characters}\marg{readings} always takes priority, and
  it ends the run.
\end{enumerate}

\noindent
Traditional characters have several accepted shapes (為 and 爲, 裡 and 裏, 說
and 説, 衛 and 衞, 線 and 綫, 溫 and 温 \dots). For word lookup these shapes
are folded together, so a word is found whichever shape is typed.

\subsection{Proofreading}

About 4\,400 characters have more than one common reading. When such a
character is not settled by a word or by a setting of your own, its reading
is only a guess. The \texttt{multiple} option formats these guesses, while
the \texttt{debug} option writes every run to the log together with the other
readings. For example,
\begin{example}
\xjyutpingsetup{multiple=\color{red}, debug}
\begin{jyutpingscope}
佢重未嚟。
\end{jyutpingscope}
\end{example}
\begin{jyutpingscope}[multiple=\color{red}]
佢重未嚟。
\end{jyutpingscope}

\noindent
The log then contains a line such as,
\begin{Verbatim}[fontsize=\small]
xjyutping> 佢keoi5:s |重cung5:m(zung6 cung4) |未mei6:s |嚟lai4:m(lei4) |
\end{Verbatim}
Each character is followed by its reading and a type, which is \texttt{w}
for a reading from the word list, \texttt{u} for a reading you set,
\texttt{m} for a guessed polyphone (with the other readings in brackets) and
\texttt{s} for a character with a single common reading. Looking at this
line, 重 means `still' here, so \verb|\setjyutping{重未}{zung6 mei6}| fixes
it, while \verb|\setjyutping{重}{zung6}| would make \emph{zung6} the default
for 重 on its own.

\section{Options}\label{sec:options}

Options can be given to \cs{usepackage}, to \cs{xjyutpingsetup} or in the
optional argument of \texttt{jyutpingscope} and \cs{xjyutping}. The options
are,

\medskip
\noindent
\begin{tabular}{@{}lll@{}}
\toprule
Key & Default & Meaning \\
\midrule
\texttt{ratio} & \texttt{0.45} & size of the Jyutping relative to the text \\
\texttt{vsep} & \texttt{1.05em} & height of the Jyutping baseline above the character's \\
\texttt{hsep} & \texttt{0.15em plus 0.4em} & space between cells, which is the least gap \\
 & & between two Jyutping plus stretch for justified lines \\
\texttt{width} & \texttt{auto} & cell width, which is \texttt{auto}, \texttt{natural} or a length \\
\texttt{font} & \cs{normalfont} & font of the Jyutping (font commands only) \\
\texttt{format} & & extra formatting, such as \verb|\color{gray}| \\
\texttt{multiple} & & formatting for guessed polyphones \\
\texttt{fancy} & \texttt{false} & tones as pitch strokes (Section~\ref{sec:fancy}) \\
\texttt{linebreak} & \texttt{false} & keep the lines of the source (Section~\ref{sec:linebreak}) \\
\texttt{align} & \texttt{justify} & alignment of the lines of a scope, which is \\
 & & \texttt{justify}, \texttt{left}, \texttt{centre} (or \texttt{center}) or \texttt{right}, \\
 & & also given alone, as in \verb|[linebreak,centre]| \\
\texttt{debug} & \texttt{false} & log every run \\
\bottomrule
\end{tabular}

\subsection{Cell width}

With \texttt{width=auto}, every character in a block gets the same cell,
which is wide enough for the longest Jyutping in that block. A length gives a
fixed grid across the whole document, where a Jyutping longer than the cell
widens its own cell rather than overlapping. On the other hand,
\texttt{width=natural} makes each cell only as wide as its own character or
Jyutping, which is tighter but uneven. For example,
\begin{example}
\xjyutping*[width=auto]{香港中文大學}
\xjyutping*[width=natural]{香港中文大學}
\xjyutping*[width=2em]{香港中文大學}
\end{example}
\begin{jyutpingscope}
\xjyutping*[width=auto]{香港中文大學}\par
\xjyutping*[width=natural]{香港中文大學}\par
\xjyutping*[width=2em]{香港中文大學}
\end{jyutpingscope}

\subsection{Size and font}

The size, font and colour of the Jyutping are set with \texttt{ratio},
\texttt{font} and \texttt{format}. For example,
\begin{example}
\xjyutping*[ratio=0.6, font=\sffamily, format=\color{gray}]{早晨，食咗飯未？}
\end{example}
\begin{jyutpingscope}
\xjyutping*[ratio=0.6, font=\sffamily, format=\color{gray}]{早晨，食咗飯未？}
\end{jyutpingscope}

\section{Fancy tones}\label{sec:fancy}

With the \texttt{fancy} option (\verb|\usepackage[fancy]{xjyutping}|), each
tone number is shown in the same way as in Visual Jyutping. Specifically, it
becomes a small stroke that traces the pitch of the tone, followed by a
smaller tone number, which is raised for the two high tones and lowered for
the others. The strokes follow the symbols of Visual Jyutping, where tone~1
is high, tone~2 rises to high, tone~3 is mid, tone~4 falls to low, tone~5
rises from low and tone~6 is low. For example,
\begin{example}
{\Large\xjyutping*[fancy]{詩史試時市事}}
\begin{jyutpingscope}[fancy]
我哋去銀行，行路返屋企。校長喺學校長大。
\end{jyutpingscope}
\end{example}
\begin{jyutpingscope}[fancy]
{\Large\xjyutping*{詩史試時市事}}\par
我哋去銀行，行路返屋企。校長喺學校長大。
\end{jyutpingscope}

\medskip\noindent
Since the strokes are drawn rather than taken from a font, they work with any
\texttt{font} and take the colour of \texttt{format} and \texttt{multiple}.
The spacing is also adjusted to the new shape. Specifically, the cells are
measured with the strokes and small numbers in place, so \texttt{width=auto}
cells come out a little wider, and the line spacing makes room for the
raised numbers. Further, if the lowered numbers reach further down than the
letters of the chosen font, the Jyutping is lifted by the difference so that
it keeps its distance from the character. As with every option,
\texttt{fancy} can be switched on or off for one scope, as in
\verb|\begin{jyutpingscope}[fancy=false]|. A syllable given without a tone
number, such as \verb|\xjyutping{唔}{m}|, is printed as it is.

\section{Verse and lyrics}\label{sec:linebreak}

The \texttt{linebreak} option keeps the lines of the source, which suits
verse and lyrics. Specifically, each line end inside \texttt{jyutpingscope}
becomes a line break, as if it were written \verb|\\|, while each blank line
leaves an empty line and starts a new paragraph. The new paragraph follows
the usual rules of the document, so it is indented by default and not
indented under \verb|\usepackage[parfill]{parskip}|, whose paragraph space
is added to the empty line. The \texttt{align} option sets the lines flush
left, centred or flush right, and a \cs{centering} inside the scope also
works. For example,
\begin{example}
\begin{jyutpingscope}[linebreak,centre]
床前明月光，
疑是地上霜。

舉頭望明月，
低頭思故鄉。
\end{jyutpingscope}
\end{example}
\begin{jyutpingscope}[linebreak,centre]
床前明月光，
疑是地上霜。

舉頭望明月，
低頭思故鄉。
\end{jyutpingscope}

\medskip\noindent
An aligned scope starts a paragraph of its own. Line ends at the start and
the end of the scope, after a \verb|\\| and before \cs{begin}, \cs{end} or a
new paragraph add no break, and a line ending in \verb|%| joins the next one
as usual. Since a line break also ends a run of characters, no word is
looked up across two lines. However, the option only works in the
environment, and only where the line ends are still in the source when the
scope begins. Hence it has no effect in \cs{xjyutping*}, in a scope inside
the argument of a command such as \verb|\parbox{...}| or in a scope inside
another scope without the option, since their text has already been read.
Inside an environment such as \texttt{minipage}, on the other hand, it
works.

\section{Xe\LaTeX{} and Lua\LaTeX{}}\label{sec:engines}

The readings, the options and the layout are the same under both engines,
while the machinery differs. Under Xe\LaTeX{}, \pkg{xeCJK} hands each
Chinese character to a hook in which \pkg{xjyutping} builds the cell of the
character. Under Lua\LaTeX{}, the rubies are built as the text is read, and
a Lua function in \texttt{xjyutping.lua}, which must be installed next to
\texttt{xjyutping.sty}, puts each cell together after \pkg{LuaTeX-ja} has
laid out the line. This gives four differences under Lua\LaTeX{},
\begin{itemize}
\item the rules of \pkg{LuaTeX-ja} for punctuation widths and line breaks
  apply, so for example a line never breaks just before a dash (——) or an
  ellipsis (……),
\item a paragraph can be of any length, and a paragraph of 40\,000
  characters compiles,
\item compiling takes about 1.6 times as long as with Xe\LaTeX{} and
\item text that comes from a macro is annotated with the options and the
  size of the place where it stands, but with the \cs{setjyutping} readings
  in force when its paragraph ends.
\end{itemize}

\section{Troubleshooting}\label{sec:limits}

\subsection{\texorpdfstring{\cs{verb}}{verb} inside a scope gives an error}

The body of \texttt{jyutpingscope} and the argument of \cs{xjyutping*} are
read in full before anything is typeset, in the same way as any argument.
Hence \cs{verb} and verbatim environments cannot go inside, and in a
\cs{url} or \cs{href} inside a scope a \verb|%| must be written \verb|\%|.
Put verbatim material outside the scope.

\subsection{Text is annotated one character at a time}

Text that comes from a macro, an \cs{include}d file or \cs{maketitle} is
annotated one character at a time, without word context, and the debug log
marks it \texttt{[no context]}. This is due to the package only seeing the
name of a macro when it reads a scope. The same applies to a file read with
\cs{input}\marg{file} inside a scope when the file uses \cs{endinput},
\cs{verb}, verbatim environments, \cs{makeatletter} or catcode changes, since
such a file is input normally instead of being read as part of the scope.
Put \cs{xjyutping*} inside the macro, or a scope inside the included file.

\subsection{A command's argument is annotated or left plain}

The arguments of \cs{label}, \cs{ref}, \cs{cite}, \cs{index} and \cs{url},
the first argument of \cs{href}, the optional argument of \cs{hyperref},
\cs{hyperlink}, \cs{pdfbookmark}, \cs{includegraphics}, the \pkg{cleveref}
commands and environment names are passed through untouched. Other commands
need their Chinese argument in braces, as in \verb|\textbf{行}| rather than
\verb|\textbf 行|.

\subsection{The table of contents has no Jyutping}

Section titles, captions and footnotes inside a scope are annotated, while
the table of contents is only annotated if \cs{tableofcontents} is itself
inside a scope, although an explicit \cs{xjyutping} in a title is annotated
there as well. Running heads and PDF bookmarks are always plain.

\subsection{A beamer frame title has no Jyutping}

In \pkg{beamer}, a \cs{frametitle} inside a scope is typeset after the scope
has ended, so it stays plain. Write \verb|\frametitle{\xjyutping*{...}}|
instead.

\subsection{Underlines break the spacing}

The underline and emphasis-mark commands of \pkg{xeCJKfntef} (Xe\LaTeX{}
only) do not keep the cell spacing inside a scope. Use \cs{underline} or
\cs{uline} from \pkg{ulem} instead.

\subsection{A long paragraph exceeds the memory of \TeX}

A scope can hold well over 100\,000 characters. However, under Xe\LaTeX{}
one paragraph is limited by the memory of \TeX{} to about 15\,000
characters, or about 13\,000 with \texttt{fancy}. Split the text into
shorter paragraphs, or compile with Lua\LaTeX{}, which has no such limit.

\subsection{A \texorpdfstring{\cs{setjyutping}}{setjyutping} in unused code still applies}

A \cs{setjyutping} inside a scope is noticed while the text is read, so one
inside \cs{iffalse}\dots\cs{fi} or an unused macro definition still applies
to the text after it. Keep conditional settings outside the scope.

\subsection{A character has no Jyutping}

Characters missing from the data are typeset in their cell without Jyutping.
Give such a character a reading with \cs{setjyutping} or \cs{xjyutping},
which both work for characters outside the data.

\subsection{pdf\LaTeX{} stops with an error}

Only Xe\LaTeX{} and Lua\LaTeX{} are supported, while pdf\LaTeX{} is not.
Compile with \texttt{xelatex} or \texttt{lualatex} instead.

\section{Data and licences}

The readings come from five groups of sources, which
\texttt{tools/fetch-sources.sh} downloads at pinned commits,
\begin{itemize}
\item the \emph{Cantonese Pronunciation List of Characters for Computers}
  (粵拼表) of the LSHK and the \pkg{rime-cantonese} dictionaries (both
  CC~BY~4.0), which give the character readings, the defaults and most of
  the words, with rime-cantonese treated as authoritative,
\item the word list of \pkg{ToJyutping} (BSD 2-Clause), which chooses
  between the readings that rime-cantonese gives a word, following Hong Kong
  usage (公園 \emph{gung1 jyun2}), and adds 512 words,
\item CC-Canto and the Cantonese readings of CC-CEDICT (both CC~BY-SA~3.0, by
  Pleco), as distributed with Jyut Dictionary, which add 2\,531 words, each
  only where it changes a reading and passes checks against rime-cantonese,
\item the book data of 粵音資料集叢, which gives the readings of 640 rare
  characters that the others lack and
\item the variant tables of OpenCC (Apache-2.0).
\end{itemize}
The word frequencies come from the list of rime-cantonese and from a count of
how often each word is used in the 6.8 million transcribed utterances of
WenetSpeech-Yue, which is kept in \texttt{tools/wenetspeech-yue-counts.tsv}.
The
tables at the top of \texttt{tools/build-data.py} hold the hand-checked
corrections, and the script regenerates \texttt{xjyutping-chars.def} and
\texttt{xjyutping-words.def} from these sources.

We measured the accuracy on seven corpora. On the Hong Kong Cantonese Corpus
(HKCanCor), whose 161\,045 characters of conversation were annotated with
Jyutping by hand, the package reads 95.6\% of the characters correctly on
the half of the files kept out of the tuning, and 97.9\% of the characters
outside sentence-final particles and interjections. On the particles of
CantoMap, which were transcribed by ear, it reads 97.5\%, and on fresh
sentences of SpiCE, MagicHub and WenetSpeech-Yue, where the systems
disagree, its reading was judged right in 91.7\% of the cases.

The package code (\texttt{xjyutping.sty} and \texttt{xjyutping.lua}) is
released under the \LaTeX{} Project Public License (LPPL) 1.3c. The data
files are released under CC~BY-SA~4.0, since they adapt the CC~BY-SA word
lists. The word list of \pkg{ToJyutping} is used under the BSD 2-Clause
License, whose notice is reproduced in \texttt{LICENSE}. The readings taken
from 粵音資料集叢 come from data published without a licence statement and
are used with attribution. The word counts of
\texttt{tools/wenetspeech-yue-counts.tsv} are counted from the transcripts of
WenetSpeech-Yue (CC~BY-NC~4.0), and no text of any corpus is included.

\section{Acknowledgements}

This package was inspired by the authors of the \LaTeX{} package
\pkg{xpinyin} by Qing Lee (李清), which puts Hanyu Pinyin above Simplified
Chinese characters, and it follows the way \pkg{xpinyin} annotates
characters through the \cs{CJKsymbol} hook of \pkg{xeCJK}. The
\texttt{fancy} option was inspired by Visual Jyutping by Vincent Tam, whose
tone symbols it draws, and by the Visual Cantonese Fonts (粵語字體) by Jon
Chui / A3I Ltd.\ (canto.hk, docs.visual-fonts.com).

We would like to thank the authors of every source of the readings,
\begin{itemize}
\item the Jyutping Workgroup of the LSHK, together with Prof Lu Qin, Dr Cheung
  Kwan Hin and Nathan Hammond, whom it thanks,
\item the Cantonese Computational Linguistics Infrastructure Development
  Workgroup (CanCLID) and the contributors of rime-cantonese and
  ToJyutping,
\item 石見田 for 粵音資料集叢, together with the authors and editors of the
  dictionaries digitised there,
\item Aaron Tan for Jyut Dictionary,
\item Pleco for CC-Canto and the Cantonese readings of CC-CEDICT, together
  with MDBG and the contributors of CC-CEDICT,
\item Carbo Kuo (BYVoid) and the contributors of OpenCC,
\item Kang Kwong Luke for the Hong Kong Cantonese Corpus (CC~BY~4.0), as
  distributed with PyCantonese,
\item Grégoire Winterstein, Carmen Tang and Regine Lai for CantoMap,
\item Khia A. Johnson, Molly Babel, Ivan Fong and Nancy Yiu for SpiCE,
\item the ASLP-lab for WenetSpeech-Yue, whose transcripts also give the word
  frequencies,
\item Beijing Magic Data Technology for the Guangzhou Cantonese
  Conversational Speech Corpus of MagicHub and
\item the contributors of the Cantonese Wikipedia, whose corpora we used to
  tune the package and to measure its accuracy.
\end{itemize}

\end{document}