The aeneas
forced aligner

Alberto Pettarin
alberto@readbeyond.it

TextAV, New York City, 2017-07-19
https://www.readbeyond.it/textav/

Summary

• https://www.readbeyond.it/aeneas/
• Forced aligner: text (fragments) + audio => time labels
• 10+ output formats (SRT, TTML, VTT, etc.)
• 30+ tested languages (ENG, ITA, SPA, etc.)
• Free software/open source (AGPLv3), code on GitHub
• Published on July 2015, now v1.7.3, estimated 1k users
• Python 2/3 (+ C) library, CLI tools, free-to-use Web App

aeneas CLI Tools

aeneasweb.org

Design Choices

• Support "marginal" languages => MFCC+DTW algo
• Easy to install ($ pip install aeneas)
• Easy to (CLI) use ($ python align.py audio text [options] output)
• User-defined granularity; designed for phrase-level (Audio-eBooks), word-level possible
• Fast: RTF <= 0.050 (see benchmark)
• RAM efficient: 4GB => up to 2h audio, 16GB => 10h, etc.

Use Cases

• Academia/Research (?)
• Audio-eBooks (ebook + audiobook, in sync)
  • ReadBeyond Free EPUB 3 Audio-eBooks
  • SIL Reading App Builder
  • Readiance.org
• Closed Captions/Timed Transcripts
  • aeneasweb.org
  • TED/TEDx Open Translation Project, Italian chapter
  • A few other a11y/public organizations/individuals

Future Work

aeneasweb.org "v1": open source (MIT), API + frontend, save projects, CC editor, lachesis lib, etc. Expected in Fall 2017.

aeneas v2: better CLI, modular functional topology, out-of-memory computation, maybe other aligner cores. No timeline yet.

Thank You!

• albertopettarin.it
• alberto@readbeyond.it
• readbeyond.it/aeneas
• aeneasweb.org