aeneas is a Python library and a set of tools to automagically synchronize audio and text.

• Version: 1.0.0
• Date: 2015-05-12
• Developed by: ReadBeyond
• Lead Developer: Alberto Pettarin
• License: the GNU Affero General Public License Version 3 (AGPL v3)
• Contact: aeneas AT readbeyond DOT it

aeneas automatically generates a synchronization map between a list of text fragments and an audio file containing the narration of the (same) text.

For example, given this text file and this audio file, aeneas computes the following map:

[00:00:00.000, 00:00:02.680] <=> 1                                                      
[00:00:02.680, 00:00:05.480] <=> From fairest creatures we desire increase,            
[00:00:05.480, 00:00:08.640] <=> That thereby beauty's rose might never die,           
[00:00:08.640, 00:00:11.960] <=> But as the riper should by time decease,              
[00:00:11.960, 00:00:15.279] <=> His tender heir might bear his memory:                
[00:00:15.279, 00:00:18.519] <=> But thou contracted to thine own bright eyes,         
[00:00:18.519, 00:00:22.760] <=> Feed'st thy light's flame with self-substantial fuel, 
[00:00:22.760, 00:00:25.719] <=> Making a famine where abundance lies,                 
[00:00:25.719, 00:00:31.239] <=> Thy self thy foe, to thy sweet self too cruel:        
[00:00:31.239, 00:00:34.280] <=> Thou that art now the world's fresh ornament,         
[00:00:34.280, 00:00:36.960] <=> And only herald to the gaudy spring,                  
[00:00:36.960, 00:00:40.640] <=> Within thine own bud buriest thy content,             
[00:00:40.640, 00:00:43.600] <=> And tender churl mak'st waste in niggarding:          
[00:00:43.600, 00:00:48.000] <=> Pity the world, or else this glutton be,              
[00:00:48.000, 00:00:53.280] <=> To eat the world's due, by the grave and thee.  

Moreover, the map can be output in several formats: SMIL for EPUB 3, SRT/TTML/VTT for closed captioning, JS for Web usage, or raw CSV/TXT/XML for further processing.

1. 2 GB RAM (4 GB recommended), 2 GHz CPU (3 GHz 64bit recommended)
2. Python 2.7.x
3. Python optional modules BeautifulSoup, numpy, and scikits (pip install ...)
4. ffmpeg and ffprobe executables installed (apt-get install ffmpeg*)
5. espeak executable installed (apt-get install espeak*)

Depending on the format(s) of audio files you work with, you might need to install additional audio codecs for ffmpeg. Similarly, you might need to install additional voices for espeak, depending on the language(s) you work on. (Installing all the codecs and all the voices available in the Debian repository might be a good idea.)

aeneas has been developed and tested on Debian 64bit, which is the only supported OS at the moment. Other Linux distributions should be good too.

However, it should work on Mac OS X and Windows as well, once you make sure ffmpeg, ffprobe and espeak are properly installed and callable by the subprocess Python module. A way to ensure the latter consists in adding the three executables to your $PATH. Alternatively, you can use VirtualBox to run aeneas inside a virtualized Debian image.

$ git clone https://github.com/readbeyond/aeneas.git
$ cd aeneas
$ pip install -r requirements.txt
$ python check_dependencies.py

If the last command prints a success message, you have all the required dependencies installed and you can confidently run aeneas in production.

1. Clone this GitHub repo:

   $ git clone https://github.com/readbeyond/aeneas.git
   $ cd aeneas

2. To compute a SMIL synchronization map map.smil for a pair (audio.mp3, text.txt), you can run:

   $ python -m aeneas.tools.execute_task audio.mp3 text.txt config_string map.smil 

   config_string is string containing all the parameters to parse text.txt correctly and to format map.smil as desired. See the documentation for details.

3. If you have several tasks to run, you can create a job container and a configuration file, and run them all at once:

   $ python -m aeneas.tools.execute_job job.zip /tmp/

   File job.zip should contain a config.txt or config.xml configuration file, providing aeneas with all the information needed to parse the input assets and format the output sync map files. See the documentation for details.

You might want to run the above modules without arguments to get their manual:

$ python -m aeneas.tools.execute_task
$ python -m aeneas.tools.execute_job

See the documentation for an introduction to the concepts of task and job, and for a list of the available options.

Online: http://www.readbeyond.it/aeneas/docs/

Generated from the source (requires sphinx):

$ git clone https://github.com/readbeyond/aeneas.git
$ cd aeneas/docs
$ make html

• Input text files in plain, parsed or unparsed format
• Text extraction from XML (e.g., XHTML) files using id and class attributes
• Arbitrary text fragment granularity (single word, subphrase, phrase, paragraph, etc.)
• Input audio file formats: all those supported by ffmpeg
• Batch processing
• Output sync map formats: CSV, JS, SMIL, TTML, TXT, VTT, XML
• Supported (= tested) languages: BG, CA, CY, DA, DE, EL, EN, ES, ET, FI, FR, GA, GRC, HR, HU, IS, IT, LA, LT, LV, NL, NO, RO, RU, PL, PT, SK, SR, SV, TR, UK
• Robust against misspelled/mispronounced words, local rearrangements of words, background noise/sporadic spikes
• Code suitable for a Web app deployment (e.g., on-demand AWS instances)

• Audio should match the text: large portions of spurious text or audio might produce a wrong sync map
• Audio is assumed to be spoken: not suitable/YMMV for song captioning
• Offline (i.e., not real time/near real time) approach
• DTW computation is memory hungry
• No protection against memory trashing

• Improving the speed of the code, especially when Sakoe-Chiba kicks in
• Improving robustness against music in background
• Isolate non-speech intervals (music, prolonged silence)
• Automated text fragmentation based on audio analysis
• Auto-tuning DTW parameters
• Reporting the alignment score
• Improving (removing?) dependency from espeak, ffmpeg, ffprobe executables
• Multilevel sync map granularity (e.g., multilevel SMIL output)
• Enforcing a max char/second constraint for CC applications
• Supporting input text encodings other than UTF-8
• Adding more languages
• Better documentation
• Testing other approaches, like HMM
• Publishing the package on PyPI

Math.

A good deal of math and computer science, a handful of software engineering and some optimization tricks.

Using the Sakoe-Chiba Band Dynamic Time Warping (DTW) algorithm to align the Mel-frequency cepstral coefficients (MFCCs) representation of the given (real) audio wave and the audio wave obtained by synthesizing the text fragments with a TTS engine, eventually mapping the computed alignment back onto the (real) time domain.

To be written. Eventually. Some day.

aeneas is released under the terms of the GNU Affero General Public License Version 3. See the LICENSE file for details.

The code for computing the MFCCs aeneas/mfcc.py is a verbatim copy from the CMU Sphinx III project.

Audio files contained in the unit tests aeneas/tests/res/ directory are adapted from recordings produced by the LibriVox Project and they are in the public domain.

Text files contained in the unit tests aeneas/tests/res/ directory are adapted from files produced by the Project Gutenberg and they are in the public domain.

No copy rights were harmed in the making of this project.

Would you like supporting the development of aeneas?

We are open to accept sponsorships to

• fix bugs,
• add new features,
• improve the quality and the performance of the code,
• port the code to other languages/platforms,
• support of third party installations, and
• improve the documentation.

Feel free to get in touch.

If you are able to contribute code directly, that's great! Feel free to open a pull request, we will be glad to have a look at it.

Please make your code consistent with the existing code base style (see the Google Python Style Guide), and test your contributed code against the unit tests before opening the pull request. Ideally, add some unit tests on the code written by you.

Please note that, by opening a pull request, you automatically agree to apply the AGPL v3 license to the code you contribute.

If you think you found a bug, please use the GitHub issue tracker to file a bug report.

Early 2012: Nicola Montecchio and Alberto Pettarin co-developed an initial experimental package to align audio and text, intended to be run locally to compute Media Overlay (SMIL) files for EPUB 3 Audio-eBooks

Late 2012-June 2013: Alberto Pettarin continued engineering and tuning the alignment tool, making it faster and memory efficient, writing the I/O functions for batch processing of multiple audio/text pairs, and started producing the first EPUB 3 Audio-eBooks with Media Overlays (SMIL files) computed automatically by this package

July 2013: incorporation of ReadBeyond Srl

July 2013-March 2014: development of ReadBeyond Sync, a SaaS version of this package, exposing the alignment function via APIs and a Web application

March 2014: launch of ReadBeyond Sync beta

April 2015: ReadBeyond Sync beta ended

May 2015: release of this package on GitHub

Many thanks to Nicola Montecchio, who suggested using MFCCs and DTW, and co-developed the first experimental code for aligning audio and text.

Paolo Bertasi, who developed the APIs and Web application for ReadBeyond Sync, helped shaping the structure of this package for its asynchronous usage.