fasrt 0.5.0

A blazing fast, zero-copy subtitle parser and writer for SRT, WebVTT and ASS/SSA in Rust.

Installation

[dependencies]
fasrt = "0.5"

Features

• Zero-copy, zero-allocation parsing — borrows directly from the input string
• #![no_std] support (with optional alloc and std features)
• Lazy iterator-based parsing — blocks are yielded on demand
• DFA-based lexing via logos for fast tokenization
• Strongly-typed newtypes (Hour, Minute, Second, Millisecond, Centisecond, Percentage) with compile-time validation
• W3C WebVTT spec conformant — validated against Web Platform Tests

SRT

SubRip has no formal specification; the cue-body behaviour below is the common ground between FFmpeg's subrip decoder (libavcodec/htmlsubtitles.c, which is also what mpv and everything else built on libavcodec uses), VLC's subtitle decoder and Aegisub's SRT reader, and each rule is documented against them.

• Strict and lossy parsing modes
• Timestamps (HH:MM:SS,mmm)
• Multiline cue bodies
• Writer (std feature)
• Cue body text parsing — two-layer design, both usable standalone on an embedded S_TEXT/UTF8 packet:
  • TextParser: logos DFA-backed, zero-alloc token stream (no_std)
  • PlainText: clean-text extraction with OnceCell-cached lazy normalization, plus an allocation-free segments() iterator
  • Tags: <b>, <i>, <u>, <s> and <font> with its attributes readable, matched case-insensitively; <br> in every form is a line break
  • A < that begins none of them is literal text, and so is the rest of the line — I <3 this and the Japanese narration convention <セリフ are subtitle text, not markup
  • Not HTML: character references are not decoded, because no SubRip reader decodes them
  • SSA ({\an8}) and MicroDVD ({Y:i}) inline codes left behind by a converter are dropped, and a brace that opens neither is text
  • Flat by construction — no tree, so no nesting depth can overflow the stack and there is no depth bound to configure

WebVTT

• WEBVTT signature and header text
• Timestamps (short MM:SS.mmm and long HH:MM:SS.mmm forms, unbounded hours)
• Cue identifiers (zero-copy &str)
• Cue settings (vertical, line, position, size, align, region)
• NOTE, STYLE, REGION blocks
• Full REGION definition parsing (id, width, lines, regionanchor, viewportanchor, scroll)
• Float percentages (e.g., 50.5%)
• CRLF, CR, LF line endings
• BOM handling
• Error recovery (--> in cue body, malformed timing lines)
• Writer with round-trip fidelity (std feature)
• Cue text parsing — two-layer design:
  • CueParser: logos DFA-backed, zero-alloc token stream (no_std)
  • CueText: W3C spec-compliant DOM tree builder with Node/TagNode types (alloc/std)
  • Tags: <b>, <i>, <u>, <c>, <ruby>, <rt>, <v>, <lang>, with classes and annotations
  • W3C tree building algorithm (§6.4): <rt> scoped on the current node, </ruby> closing an open <rt> along with its <ruby>, unterminated tag handling
  • §6.4's list of applicable classes (TagNode::classes), which excludes the empty classes between adjacent . separators — TagNode::classes_raw keeps the source text
  • §6.4's applicable language, derived from the tree rather than stored on every node: TagNode::declared_language for one step of a descent, CueText::nodes_with_language for the whole tree in document order
  • Bounded nesting depth (cue::Options::max_depth, default 16), so a hostile cue cannot overflow the stack while the tree is built, walked or dropped — every walk at the default limit is held to a 128 KiB thread by the test suite; CueText::try_parse refuses such input rather than flattening it
  • Full HTML5 named character reference support (2,231 entities via phf perfect hash map)
  • Numeric ( ) and hexadecimal ( ) character references
  • Lazy text normalization via CueStr with OnceCell-cached decoding and NULL (U+0000 → U+FFFD) replacement

ASS/SSA

ASS/SSA has no formal specification; the behaviour below follows the SSA v4.00 specification document together with the VSFilter/libass renderers.

• Sections ([Script Info], [V4 Styles], [V4+ Styles], [Events], [Fonts], [Graphics], and unknown sections verbatim)
• Timestamps (H:MM:SS.cc, centiseconds, unpadded and unbounded hours)
• Event rows (Dialogue:, Comment:, Picture:, Sound:, Movie:, Command:)
• Format:-driven field order — EventFormat presets for ASS v4+, SSA v4 and Matroska packets
• The Name (a.k.a. Actor) speaker column surfaced as first-class data
• Style: rows exposed as zero-copy fields (style rendering semantics are out of scope)
• [Fonts] / [Graphics] payload lines preserved verbatim, never mistaken for properties
• Strict and lossy parsing modes
• CRLF, CR, LF line endings, and BOM handling
• Writer with round-trip fidelity (std feature)
• Event text parsing — two-layer design, both usable standalone:
  • TextParser: logos DFA-backed, zero-alloc token stream (no_std, no alloc needed)
  • PlainText: clean-text extraction with OnceCell-cached lazy normalization, plus an allocation-free segments() iterator
  • Override blocks {…}, with libass tag boundaries: longest-match names (\fscx is not \fs + cx), spaces skipped after the backslash, and argument lists ending at the first ) (\t(0,500,\frz360) stays one tag)
  • libass brace semantics: { opens a block only when a } follows, \{ and \} are literal-brace escapes, so cleaning never deletes text a renderer would show
  • \N, \n and \h escapes; \p<n> drawing mode is tokenized and skipped, never interpreted as geometry
  • Columns a Format: line declares under an unrecognized name are preserved through a write, not dropped
• Standalone event parsing for embedded Matroska S_TEXT/ASS tracks, which arrive one event per packet with container timing authoritative

Optional dependencies

Benchmarks

Measured on Apple Silicon with cargo bench (Criterion).

SRT

WebVTT

Cue Text

Run benchmarks yourself:

cargo bench

License

fasrt is under the terms of both the MIT license and the Apache License (Version 2.0).

See LICENSE-APACHE, LICENSE-MIT for details.

Copyright (c) 2026 FinDIT Studio authors.