Coding With Fun
Home Docker Django Node.js Articles FAQ

Web Chinese font processing summary

May 30, 2021 Article blog

Table of contents


In a Web project, using the right font can provide a good experience for users. B ut there are too many font files, and if you want to see the font effect, you can only open it one by one, which greatly affects your productivity. T herefore, you need to implement a feature that allows you to preview fonts based on fixed text and user input. There are two main issues that are addressed in the implementation of this function:

  • Chinese font size is too large and the load time is too long
  • The preview is not displayed until the font load is complete

Now summarize the solution of the problem and my thoughts.

 Web Chinese font processing summary1

Use the web to customize fonts

Before we talk about these two issues, let's briefly describe how to use a web custom font. T o use a custom font, you can rely on @font-face rules defined by CSS Fonts Module Level 3. Here's a basic way to use it that's compatible with all browsers:

@font-face {
    font-family: "webfontFamily"; /* 名字任意取 */
    src: url('webfont.eot');
         url('web.eot?#iefix') format("embedded-opentype"),
         url("webfont.woff2") format("woff2"),
         url("webfont.woff") format("woff"),
         url("webfont.ttf") format("truetype");
.webfont {
    font-family: webfontFamily;   /* @font-face里定义的名字 */

Because woff2 woff and ttf formats are already well supported in most browsers, the code above can also be written as:

@font-face {
    font-family: "webfontFamily"; /* 名字任意取 */
    src: url("webfont.woff2") format("woff2"),
         url("webfont.woff") format("woff"),
         url("webfont.ttf") format("truetype");

With @font-face rule, we simply upload the font source file to cdn, have the url value of @font-face rule as the address of the font, and finally apply this rule to the web text to preview the font.

But in doing so, we can clearly see a problem, font size is too large resulting in too long load time. We open the browser's Network panel to view:

 Web Chinese font processing summary2

You can see that the font has a volume of 5.5 MB and a load time of 5.13 s. W hile many of the quark platforms Chinese font sizes between 20 and 40 MB, it can be expected that the load time will increase further. This wait time is not acceptable if the user is still in a weak network environment.

1. Chinese font size is too large and the load time is too long

1. Analysis of the reasons

So why is the Chinese font so large compared to the size of the English font, which is mainly due to two reasons:

  1. Chinese font contains a large number of glyphs, while English fonts contain only 26 letters and some other symbols.
  2. Chinese glyph lines are far more complex than English lines, and are used to control the position points of Chinese glyph lines more than English lines, so the amount of data is greater.

With opentype.js we can count the difference between the number of glyphs and the number of glyphs in a Chinese font and an English font:

The font name The number of glyphs The number of bytes held by glyphs
FZQingFSJW_Cu.ttf 8731 4762272
JDZhengHT-Bold.ttf 122 18328

Quark platform font previews need to be done in two ways, one is a fixed character preview, and the other is based on characters entered by the user. But either way, only a small number of characters will be used for the font, so it is not necessary to load the font in full, so we need to streamline the font file.

2. How to reduce the size of a font file


The unicode-range property is typically used in conjunction with @font-face rule, which controls the use of specific fonts for specific characters. But it doesn't reduce the size of the font file, and interested readers can try it.


fontmin is a subset of fonts implemented by JavaScript As mentioned earlier, Chinese font size is larger than English font because it has more glyphs, then the idea of streamlining a font file is to remove useless glyphs:

// 伪代码
const text = '字体预览'
const unicodes = text.split('').map(str => str.charCodeAt(0))
const font = loadFont(fontPath)
font.glyf = => {
 // 根据unicodes获取对应的字形

In fact, thinning is not that simple, because a font file consists of many 表(table) that are associated with each other, such as maxp table that records the number of glyphs and the offset of the glyph position stored in the loca table. At the same time, the font file starts with offset table(偏移表) which offset table information for all tables in the font, so if we change glyf table, we'll update the other tables at the same time.

Before we discuss how fontmin performs font interception, let's look at the structure of the font file:

 Web Chinese font processing summary3

The structure above is limited to a font file that contains only one font, and the glyph outline is based on TrueType format, which determines the value of sfntVersion so the offset table starts with 0字节 of the font file. If the font file contains more than one font, the offset table for each font is specified in TTCHeader, which is outside the scope of the article.

Offset table:

Type Name Description
uint32 sfntVersion 0x00010000
uint16 numTables Number of tables
uint16 searchRange (Maximum power of 2 <= numTables) x 16.
uint16 entrySelector Log2(maximum power of 2 <= numTables).
uint16 rangeShift NumTables x 16-searchRange.

Table record:

Type Name Description
uint32 tableTag Table identifier
uint32 checkSum CheckSum for this table
uint32 offset Offset from beginning of TrueType font file
uint32 length Length of this table

For a font file, whether its glyph outline is TrueType format or PostScript-based CFF format, it must contain tables such as cmap head hhea htmx maxp name OS/2 post I f its glyph outline is trueType format, there are cvt fpgm glyf loca prep gasp six tables will be used. The six tables are optional except for glyf and loca

Fontmin intercept glyph principle

fontmin uses fonteditor-core internally, and the core font processing is left to this dependency, fonteditor-core main process is as follows:

 Web Chinese font processing summary4

1. Initialize reader

Convert font files to ArrayBuffer for subsequent reading of data.

2. Extract Table Directory

As we mentioned earlier, the structure immediately after offset table(偏移表) is table table record(表记录) and multiple table record are called Table Directory fonteditor-core reads table Table Directory of the original font first, and we know from the structure recorded in the table above that each table record has four fields, each of which accounts for 4 bytes, so it is convenient to read with DataView and finally get all the table information for a font file as follows:

 Web Chinese font processing summary5

3. Read the table data

At this step, table data is read based on the offset and length information recorded by Table Directory T he contents of glyf table are the most important for thin fonts, but glyf table record only tells us the length of glyf table and the offset of glyf table relative to the entire font file, so how do we know the number, location, and size of glyf in the glyf table? This requires specifying the number of glyphs with the maxp table and loca(glyphs location) table in the font, and the loca table records the offset of all glyphs from glyf table in the font as follows: maxp numGlyphs

Glyph Index Offset Glyph Length
0 0 100
1 100 150
2 250 0
... ... ...
n-1 1170 120
extra 1290 0

According to the specification, index 0 points to the missing (missing character) which is the character that appears when a character cannot be found in the font, which is usually represented by a blank box or space, and when the missing character does not have an outline, the loca[n] = loca[n+1] given according to the definition of the loca table. We can find that one extra has extra added to the table above to calculate the length of the last glyph loca[n-1]

The Offset field value in the table above is in bytes, but the exact number of bytes depends on the indexToLocFormat field value of the font head table, when this value is 0 Offset 100 equals 200 bytes, and when this value is 1 Offset 100 is equal to 100 bytes, two different cases that correspond to Short version Long version in the font.

But just knowing the offset of all glyphs is not enough, and we can't recognize which glyph is what we need. S uppose I need 字体预览 four glyphs, and the font file has ten thousand glyphs, and we know the offset of all glyphs through loca table, but which four blocks of data represent 字体预览 So we also need to use cmap table to determine the specific glyph location, cmap table records the character code (unicode) to the glyph index mapping, we get the corresponding glyph index, we can get the glyph offset in the glyf table based on the index.

 Web Chinese font processing summary6

The data structure of a glyph begins with Glyph Headers

Type Name Description
int16 numberOfContours the number of contours
int16 xMin Minimum x for coordinate data
int16 yMin Maximum y for coordinate data
int16 xMax Minimum x for coordinate data
int16 yMax Maximum x for coordinate data

numberOfContours field specifies the number of outlines for this glyph, and the data structure immediately after Glyph Headers is Glyph Table

In the definition of a font, the outline is made up of position points, and each location point has numbers, which start at 0 in ascending order. So the glyph we read is to read the values in Glyph Headers and the position point coordinates of the outline.

In Glyph Table an array consisting of the number of the last position point for each profile is stored, from which you can find that the glyph has a total of several position points. For example, if the value of this array is [3, 6, 9, 15] you can tell that the number of the last position point on the fourth outline is 15, then the glyph has a total of 16 position points, so we only need to traverse arrayBuffer 16 times for loop access to get the coordinate information of each location point, thus extracting the glyph we want, which is how fontmin is when intercepting glyphs.

In addition, when extracting coordinate information, the coordinate value of the other location points is not absolute except for the first location point, for example, the coordinates of the first point are [100, 100] the second read value is [200, 200] [200, 200] [300, 300]

Because a font involves so many tables, and the data structure of each table is different. It is not possible to list how fonteditor-core handles each table.

4. Correlate glyf information

In fonts that use trueType outlines, each glyph provides the values of xMin xMax yMin and yMax which are Bounding Box in the figure below. I n addition to these four values, two fields, advanceWidth and leftSideBearing which are not in the glyf table and are therefore not available when the glyph information is intercepted. In this step, fonteditor-core reads the hmtx table to get both fields.

 Web Chinese font processing summary7

5. Write the font

In this step, the size of the font file is recalculated and the values associated with 偏移表(Offset table) and table 表记录(Table record) are updated, and then the 偏移表 表记录 table 表数据 is written to the file in turn. I t is important to note that when writing 表记录 you must write by table name order. For example, if there are four tables: prep hmtx glyf head the order of writing should be glyf -> head -> hmtx -> prep which is not required for 表数据

Fontmin deficiencies

fonteditor-core only the fourteen tables mentioned above during the process of intercepting fonts, and the rest are discarded. Each font typically also contains two tables, vhea and vmtx which control information such as the spacing of the font when it is in a vertical layout, which is lost when the font is intercepted with fontmin and the difference can be seen when the text is displayed vertically (on the right is after the intercept):

 Web Chinese font processing summary8

Fontmin uses methods

Once we understand how fontmin works, we can use it happily. After the server accepts the fontmin request from the client, the font is intercepted by fontmin returns the buffer corresponding to the intercepted font file, and don't forget that the font path in @font-face rule base64 format, so we just need to embed the Buffer into base64 format and return it to the client in @font-face and then the client inserts the @font-face into <head></head> label as CSS.

For fixed previews, we can also save the font file on the CDN, but the disadvantage of this approach is that if the CDN is unstable, the font load will fail. I f, in the above method, each intercepted font exists as a base64 string, you can make a cache on the service side, and there is no problem. The font subset code generated with fontmin is as follows:

const Fontmin = require('fontmin')
const Promise = require('bluebird')

async function extractFontData (fontPath) {
  const fontmin = new Fontmin()
      text: '字体预览'

  await Promise.promisify(, { context: fontmin })()

For fixed preview content we can pre-generate split fonts, and for dynamic preview content entered by users, of course we can follow this process:

Get input -> Intercept glyph -> Upload CDN -> generate @font-face -> insert page

Following this process, clients need to request twice to get font resources (don't forget to actually request fonts until @font-face inserts a page), and there's no better way 截取字形 and 上传 CDN We know that the outline of a glyph is determined by a series of position points, so we can get the coordinates of the position points in glyf table and draw a particular glyph directly through SVG image.

SVG is a powerful image format that you can interact with using CSS and JavaScript where path element is primarily applied

Getting location information and generating path tags we can do with opentype.js and once the client gets the path element that is entered into the glyph, it just needs to traverse the generated SVG label.

3. The advantages of reducing the size of font files

A comparison table of file size and load speed after font interception is attached below. As you can see, fonts are loaded 145 times faster after being intercepted than when loaded in full.

fontmin supports the generation of woff2 files, but the official documentation is not updated, and I started with woff files, but woff2 format files are smaller and browser support is good

The font name size time
HanyiSentyWoodcut.ttf 48.2MB 17.41s
HanyiSentyWoodcut.woff 21.7KB 0.19s
HanyiSentyWoodcut.woff2 12.2KB 0.12s

Second, the font load is completed before the preview content is displayed

This is the second problem in implementing the preview function.

There are two concepts in the browser's font display behavior, 阻塞期 and 交换期 in the case of Chrome there is a period of time before the font is loaded, which is called 阻塞期 I f the load is still not completed during the 阻塞期 the fallback font is displayed first, enters the 交换期 and waits for the font to be replaced when the load is complete. T his causes the page font to flash, which doesn't match the effect I want. And the font-display property controls this behavior of the browser, can we change the value of the font-display property to achieve our purpose?


Block Period Swap Period
block Short Infinite
swap None Infinite
fallback Extremely Short Short
optional Extremely Short None

The display policy of the font is related to the value of font-display and the browser's default font-display value is auto which behaves closely to the value block

The first strategy is FOIT(Flash of Invisible Text) FOIT default representation of a browser when loading fonts, whose rules are as mentioned earlier.

The second strategy is FOUT(Flash of Unstyled Text) FOUT instructs the browser to use the fallback font until the custom font is loaded, with a value of swap

Apps with two different strategies: Google Fonts FOIT Hanyi Word Library FOUT

In a quark project, I want the effect to be that the preview is not displayed until the font is loaded, and the FOIT strategy is the closest. However, the maximum time FOIT text content is not visible is 3s and if the user's network is not in good condition, the back-up font is displayed after 3s causing the page font to blink, so font-display property does not meet the requirements.

As you can see, the CSS Font Loading API also provides solutions at the JavaScript level:


Let's first look at their compatibility:

 Web Chinese font processing summary9

 Web Chinese font processing summary10

It's IE again, and IE doesn't have a user to worry about

We can construct a FontFace object from the FontFace constructor:

const fontFace = new FontFace(family, source, descriptors)

  • family
    • Font name, specifying a name as the value of the CSS property font-family
  • source
    • The font source, which can be a url or ArrayBuffer
  • descriptors optional
    • style: font-style
    • weight: font-weight
    • stretch: font-stretch
    • display: font-display (this value can be set, but will not take effect).
    • unicodeRange: @font-face unicode-ranges
    • variant: font-variant
    • featureSettings: font-feature-settings

FontFace is not loaded after a fontFace is constructed, and fontFace load method must be executed. load method returns a promise and the resolve value of promise is the font after it has been loaded successfully. However, just loading successfully does not make this font effective, and the returned fontFace needs to be added to fontFaceSet

Here's how to use it:

  * @param {string} path 字体文件路径
async function loadFont(path) {
  const fontFaceSet = document.fonts
  const fontFace = await new FontFace('fontFamily', `url('${path}') format('woff2')`).load()

Therefore, when the client can set the CSS of the text content to opacity: 0 wait for await loadFont(path) to be executed, and then set the CSS to opacity: 1 so that we can control that the content is not displayed until the custom font load is complete.

Final summary

This article describes the problems and solutions encountered in developing font preview functionality, limited to OpenType specification entries, and in the Introduction fontmin Principles section, only describes the handling of glyf tables, which interested readers can learn more about.

This work review and summary process, but also thinking about better implementation, if you have suggestions welcome to communicate with me. At the same time the content of the article is my personal understanding, there are errors difficult to avoid, if found errors welcome to correct.

Thanks for reading!


Author: Lin Lin

Source: Bump Lab