kara-templater

An ASS karaoke template engine rewritten in Python and C++. It runs without Aegisub or Lua.

The package keeps the karaoke templater's ASS template format, $variables, modifiers, and helper functions. Expressions inside !…! and code lines use Python. Lua code in existing templates must be rewritten manually; there is no Lua compatibility interpreter.

Installation

Requires Python 3.12 or newer. Binary wheels target CPython 3.12–3.14 on Linux x86-64 (glibc 2.28+) and Windows 10+ x64. They include the native library dependencies; no compiler or MSYS2 installation is needed to use a wheel.

python -m pip install kara-templater

Install the fonts specified by your subtitle styles. Missing fonts use system font fallback, and different fonts can change the layout.

Build from source on Linux

Install a C/C++ toolchain and native dependencies.

Arch Linux:

sudo pacman -S --needed base-devel python python-pip autoconf automake libtool nasm \
  freetype2 harfbuzz fribidi fontconfig libpng

Debian / Ubuntu, with Python 3.12+:

sudo apt install build-essential python3-dev python3-venv autoconf automake \
  libtool nasm pkg-config libfreetype-dev libharfbuzz-dev libfribidi-dev \
  libfontconfig-dev libpng-dev

From the project directory:

python -m venv .venv
. .venv/bin/activate
python -m pip install .

Build from source on Windows

Use 64-bit Python from python.org and MSYS2. In the MSYS2 UCRT64 shell, install the native toolchain:

pacman -Syu
pacman -S --needed patch mingw-w64-ucrt-x86_64-gcc \
  mingw-w64-ucrt-x86_64-pkgconf mingw-w64-ucrt-x86_64-nasm \
  mingw-w64-ucrt-x86_64-freetype mingw-w64-ucrt-x86_64-harfbuzz \
  mingw-w64-ucrt-x86_64-fribidi

Restart the shell and complete the update if MSYS2 requests it. Then, in PowerShell at the project directory, build and install a wheel. Adjust the MSYS2 path if it is not installed at C:\msys64.

$env:Path = "C:\msys64\ucrt64\bin;C:\msys64\usr\bin;$env:Path"
python -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install delvewheel
python tools/collect_licenses.py
python -m pip wheel . --no-deps --wheel-dir dist
python -m delvewheel repair --strip --no-mangle libharfbuzz-0.dll --wheel-dir wheelhouse (Get-ChildItem dist\*.whl).FullName
python -m pip install --no-index --find-links wheelhouse kara-templater

The repaired wheel bundles the required DLLs so it can run without MSYS2 on the destination machine. Windows uses the system DirectWrite/GDI font provider.

Command line

kara-templater input.ass output.ass
python -m kara_templater input.ass output.ass

Try the included example from a source checkout:

kara-templater examples/karaoke.ass output.ass

To correct horizontal coordinates for the video's actual dimensions:

kara-templater input.ass output.ass --video-size 1920 1080

Templates are ASS Comment events. Put the template type and modifiers in the Effect field. For example:

• Effect: template syl noblank
• Text: !retime("syl")!{\an5\pos($scenter,$smiddle)\fad(50,80)}

For lyrics such as {\k50}你{\k50}好, this generates one positioned effect event per nonblank syllable.

The output preserves templates, turns processed lyrics into comments with Effect karaoke, and marks generated events with Effect fx. Running the engine again removes old fx events and regenerates them.

Python API

from kara_templater import Document, Style, apply_templates, text_extents

source = Document.load("input.ass")
result = apply_templates(source)
result.save("output.ass")

width, height, descent, external_leading = text_extents(
    Style(fontname="Noto Sans CJK SC", fontsize=48), "你好"
)

apply_templates returns a new document without modifying its input. Document.parse(text) and result.dumps() support in-memory use.

Templates

Modifiers include all, repeat N / loop N, notext, keeptags, noblank, multi, fx NAME, and fxgroup NAME. Templates match their own style by default; all matches every style. Unknown modifiers, variables, and execution errors report the template event number instead of being silently ignored.

• \k, \K, \kf, \ko: syllable duration in centiseconds.
• \-NAME: inline effect name, inherited by subsequent syllables.
• # / ＃: continue the preceding syllable; multi runs the template separately for each highlight.
• 漢|かん / 漢｜かん: separate base text from furigana. A furigana prefix ! starts a new group; < allows spillback to the left. Full-width prefixes are also accepted. A missing furigana style is created as STYLE-furigana at half the base font size.

Variables and execution environment

$variables are case-insensitive:

• Line timing and indices: $lstart, $lend, $ldur, $lmid, $li, $syln.
• Syllable timing and indices: $sstart, $send, $sdur, $skdur, $smid, $si.
• Layout: prefix l or s to left, center, right, width, top, middle, bottom, height, x, or y.
• Current-scope aliases: $start, $end, $dur, $kdur, $mid, $i, and unprefixed layout variables.
• Other values: $layer, $style, $actor, $margin_l, $margin_r, $margin_v, $margin_t, $margin_b.

Expressions and code share Python variables. The environment provides math, random, meta, styles, orgline, line, syl, basesyl, j, maxj, fxgroup, and text_extents. line is the current mutable output event; in a code template it is the source lyric event. syl is a dictionary with attribute access.

Example Effect and Text pairs:

code once                  → amplitude = 12
code line                  → fxgroup["spark"] = orgline.actor == "solo"
template syl repeat 3      → !retime("syl")!{\an5\pos(!$x + amplitude * j!,$y)}

An expression returning None produces an empty string. Integral floating-point values are written without a trailing .0.

Helpers

• retime(mode, addstart=0, addend=0): modes are syl, presyl, postsyl, line, preline, postline, start2syl, syl2end, sylpct, and set / abs. Offsets are milliseconds, except that sylpct uses percentages of syllable duration. Timing is based on the current line, so consecutive calls use the previously modified times.
• relayer(layer), restyle(style): modify the current output event.
• maxloop(count) / maxloops(count), loopctl(j, count): control template loops dynamically.
• remember(name, value, decorator=None), recall(name, default=None), remember_if(name, value, condition, decorator=None): store and retrieve values. The optional decorator is a Python function mapping a name to a memory key.
• remember_line, remember_syl, remember_basesyl: isolate remembered values by source event, current syllable, or base syllable.

Limitations and safety

• Only process trusted subtitles. Python expressions and code have the permissions of the current process. This is not a sandbox. Template iteration and output limits cannot stop arbitrary Python code from looping forever or performing malicious operations.
• The defaults are 10,000 iterations per template loop and 100,000 output events. The Python API accepts max_iterations and max_output_lines to change these limits.
• Only ASS v4.00+ is supported, not legacy SSA styles. Extra fields, unknown sections, and attachments are preserved, but byte-for-byte formatting is not guaranteed.
• Layout targets single-line karaoke using the base style. It does not simulate automatic wrapping, event collisions, inline style overrides, \pos / \move, or drawing geometry. text_extents supports \N hard breaks; syllable layout is not multiline-aware.
• text_extents returns logical advance width, font height, and descent, not visible glyph bounds. It preserves edge spaces and measures braces literally. External leading is always 0. Pixel-identical results across fonts, platforms, or Aegisub's native font interfaces are not guaranteed.
• No GUI, video/audio processing, Automation plugin interface, or other Aegisub features are included.