Split a single, monolithic mp3 audiobook file into chapters using Machine Learning and ffmpeg.
This is a simple command line utility that will chapterize your mp3 audiobooks for you. No longer will you have to dissect a waveform to look for chapter breaks, or deal with all the annoyances that come with a single audiobook file.
You can use this as an intermediary step for creating .m4b files, or keep the files the way they are if you don't want to sacrifice audio quality through re-encoding (or prefer .mp3 files for some reason).
The script utilizes the vosk-api machine learning library which performs a speech-to-text conversion on the audiobook file, generating timestamps throughout which are stored in a srt (subrip) file. The file is then parsed, searching for phrases like "prologue", "chapter", and "epilogue", which are used as separators for the generated chapter files.
NOTE: Downloading models requires the
requestslibrary
The small United States English (en-us) machine learning model is already provided with this project. However, various other languages and model sizes are available to download directly from the script - it will unpack the archive for you, too, and no additional effort should be required. See Supported Languages and Models for a comprehensive list.
Models should be saved within the project's model directory as this is where the script looks for them. If you want to save multiple models or sizes, no problem - just don't change the name of the unpacked model archive as the script uses portions of it to determine which model to use based on the combination of arguments passed.
If there is a model conflict, the script will perform a best effort guess to determine which one to use. If that fails, an error will be thrown.
The script will also parse metadata from the source file along with the cover art (if present) and copy it into each chapter file automatically. There are CLI parameters you can use to pass your own ID3-compliant metadata properties, too, which always take precedence over the fields extracted from the source file (if there is a conflict). Otherwise, the tags will be combined and added to the final output.
Cue files can be created to make editing chapter markers and start/stop timecodes easier. This is especially useful if the machine learning speech-to-text conversion misses a section, timecodes require tweaking, or if you wish to add additional breakpoint sections (such as 'forward' or 'aftermath') which are not currently supported by the script's parsing engine (and may not be supported anytime soon due to the difficulty required to make them work consistently without breaking other stuff).
Cue files are always generated within the same directory as the audiobook itself, but there are arguments which allow you to specify a custom path to a cue file if it is inconvenient to keep them paired with the audiobook.
Cue file syntax is generated in a somewhat unconventional way to make script parsing easier, but is similar enough that it can be converted to a standard .cue format fairly easily.
Included with this project is a defaults.toml file which you can use to specify configurations that you use frequently. This removes the need to pass certain CLI argument each time you run the script and provides options for specifying file paths (such as the path to ffmpeg if you don't feel like setting environment variables - see below).
Here is the base template included with the project:
# Uncomment lines starting with '#' to set config options, or modify an existing line.
# No spaces before/after the '='!
#
# Default model language
default_language='english'
# Default model size
default_model='small'
# Defaults to the system's PATH. You can specify a full path in single '' quotes
ffmpeg_path='ffmpeg'
# Change this to True if you always want the script to generate a cue file
generate_cue_file='False'
# Set this to the cue file path you want to use. Useful for continuous edits and script runs where the
# cue file is saved somewhere other than the current audiobook directory (the default search path).
# The cue_path script argument takes precedence over this path if used
cue_path=''To install python dependencies, open a command shell and type the following:
NOTE: If you're on Linux, you might need to use
pip3instead
# Using the requirements file (recommended)
pip install -r requirements.txt
# Manually installing packages
pip install vosk rich requestsIt is recommended that you add ffmpeg to your system PATH so you don't have to run the script from the same directory. How you do this depends on your Operating System; consult your OS documentation (if you aren't familiar with the process, it's super easy. Just Google it).
Here is a quick example for Windows using PowerShell (it can be done via GUI, too):
# Whatever the path is to your ffmpeg install
$ffmpeg = 'C:\Users\SomeUser\Software\ffmpeg.exe'
$newPath = $env:PATH + ";$ffmpeg"
[Environment]::SetEnvironmentVariable('PATH', $newPath, 'User')
# Now close and reopen PowerShell to updateHere is a quick example using bash:
# Set this equal to wherever ffmpeg is
ffmpeg="/home/someuser/software/ffmpeg"
# If you're using zsh, replace with .zshrc
echo "export PATH=${ffmpeg}:${PATH}" >> ~/.bashrc
# Source the file to update
source ~/.bashrcIf you don't want to deal with all that, you can add the path of ffmpeg to the defaults.toml file included with the repository - copy and paste the full path and set it equal to the ffmpeg_path option using single quotes '':
# Specfying the path to ffmpeg manually
ffmpeg_path='C:\Users\SomeUser\Software\ffmpeg.exe'
# If ffmpeg is added to PATH, leave the file like this
ffmpeg_path='ffmpeg'NOTE: You can set a default language and model size in the defaults.toml file included with the repository
The vosk-api provides models in several languages. By default, only the small 'en-us' model is provided with this repository, but you can download additional models in several languages using the script's --download_model/-dm parameter, which accepts arguments small and large (if nothing is passed, it defaults to small); if the model isn't English, you must also specify a language using --language/-l parameter. See Usage for more info.
Not all models are supported, but you can download additional models manually from the vosk website (and other sources listed on the site). Simply replace the existing model inside the /model directory with the one you wish to use.
The following is a list of models which can be downloaded using the --download_model parameter of the script:
You can use either the Language or Code fields to specify a model
| Language | Code | Small | Large |
|---|---|---|---|
| English (US) | en-us | ✓ | ✓ |
| English (India) | en-in | ✓ | ✓ |
| Chinese | cn | ✓ | ✓ |
| Russian | ru | ✓ | ✓ |
| French | fr | ✓ | ✓ |
| German | de | ✓ | ✓ |
| Spanish | es | ✓ | ✓ |
| Portuguese | pt | ✓ | ✓ |
| Greek | el | ✕ | ✓ |
| Turkish | tr | ✕ | ✓ |
| Vietnamese | vn | ✓ | ✕ |
| Italian | it | ✓ | ✓ |
| Dutch | nl | ✓ | ✕ |
| Catalan | ca | ✓ | ✕ |
| Arabic | ar | ✕ | ✓ |
| Farsi | fa | ✓ | ✓ |
| Filipino | tl-ph | ✕ | ✓ |
| Kazakh | kz | ✓ | ✓ |
| Japanese | ja | ✓ | ✓ |
| Ukrainian | uk | ✓ | ✓ |
| Esperanto | eo | ✓ | ✕ |
| Hindi | hi | ✓ | ✓ |
| Czech | cs | ✓ | ✕ |
| Polish | pl | ✓ | ✕ |
The model used for speech-to-text the conversion is fairly dependent on the quality of the audio. The model included in this repo is meant for small distributions on mobile systems, as it is the only one that will fit in a GitHub repository. If you aren't getting good results, you might want to consider using a larger model (if one is available).
usage: chapterize_ab.py [-h] or [--help]
usage: chapterize_ab.py [-ll] or [--list_languages]
usage: chapterize_ab.py [AUDIOBOOK_PATH] [--timecodes_file [TIMECODES_FILE]] [--language [LANGUAGE]]
[--download_model [{small,large}]] [--narrator [NARRATOR]] [--comment [COMMENT]]
[--model [{small,large}]] [--cover_art [COVER_ART_PATH]] [--author [AUTHOR]]
[--year [YEAR]] [--title [TITLE]] [--genre [GENRE]] [--write_cue_file]
[--cue_path [CUE_PATH]]
positional arguments:
AUDIOBOOK_PATH path to audiobook mp3 file. required.
optional argument flags:
-h, --help show help message with usage examples and exit.
-ll, --list_languages list supported languages and exit.
-wc, --write_cue_file generate a cue file inside the audiobook directory for editing chapter markers.
default disabled, but can be enabled permanently through defaults.toml.
optional arguments:
--timecodes_file, -tc [TIMECODES_FILE]
DESCRIPTION: optional path to an existing srt timecode file in a different directory.
--language, -l [LANGUAGE]
DESCRIPTION: model language to use. requires a supported model.
en-us is provided with the project.
--model, -m [{small,large}]
DESCRIPTION: model type to use where multiple models are available. default is small.
--download_model, -dm [{small,large}]
DESCRIPTION: download a model archive. language to download specified via
the --language argument.
--cover_art, -ca [COVER_ART_PATH]
DESCRIPTION: path to cover art file. Optional.
--author, -a [AUTHOR]
DESCRIPTION: audiobook author. Optional metadata field.
--narrator, -n [NARRATOR]
DESCRIPTION: audiobook narrator (should be compatible with most players).
optional metadata field.
--title, -t [TITLE]
DESCRIPTION: audiobook title. optional metadata field.
--genre, -g [GENRE]
DESCRIPTION: audiobook genre. optional metadata field. multiple genres can be separated
by a semicolon
--year, -y [YEAR]
DESCRIPTION: audiobook release year. optional metadata field.
--comment, -c [COMMENT]
DESCRIPTION: audiobook comment. optional metadata field.
--cue_path, -cp [CUE_PATH]
DESCRIPTION: path to cue file in non-default location (i.e., not in the audiobook directory)
containing chapter timecodes. can also be set within defaults.toml, which has
lesser precedence than this argument.
NOTE: Each argument has a shortened alias. Most examples use the full argument name for clarity, but it's often more convenient in practice to use the aliases
# Adding the title and genre metadata fields
~$ python3 ./chapterize_ab.py '/path/to/audiobook/file.mp3' --title 'Game of Thrones' --genre 'Fantasy'# Adding an external cover art file (using shorthand flag -ca)
PS > python .\chapterize_ab.py 'C:\path\to\audiobook\file.mp3' -ca 'C:\path\to\cover_art.jpg'# Set model to use German as the language (requires a different model, see above)
PS > python .\chapterize_ab.py 'C:\path\to\audiobook\file.mp3' --language 'de'# Download a different model (Italian large used here as an example)
~$ python3 ./chapterize_ab.py '/path/to/audiobook/file.mp3' --download_model 'large' --language 'italian'# Write a cue file inside the audiobook directory using the --write_cue_file option flag
PS > python .\chapterize_ab.py 'C:\path\to\audiobook\file.mp3' --write_cue_file# Specify custom path to a cue file. Overrides the default search path (audiobook directory)
~$ python3 ./chapterize_ab.py '/path/to/audiobook/file.mp3' --cue_path '/path/to/file.cue'This makes the assumption that you are familiar with building and running Docker. There is a Dockerfile that you can build and run this solution with.
Build with the following:
docker build -t chapterizeaudiobooks .With a volume setup, you can run this solution like so:
docker run --rm -v "path/on/your/machine/to/audiobooks:/audiobooks" chapterizeaudiobooks "/audiobooks/file.mp3"This script is still in alpha, and thus there are bound to be some issues; I've noticed a few words and phrases that have falsely generated chapter markers, which I'm currently compiling into an ignore list as I see them. With that said, it's been remarkably accurate so far.
I encourage anyone who might use this to report any issues you find, particularly with false positive chapter markers. The more false positives identified, the more accurate it will be!
So far, support for this project is primarily targeted toward English audiobooks only; I've added some German content, but I'm by no means a fluent speaker and there are a lot of gaps.
If you want to contribute an exclusion list and chapter markers for other languages (preferably vosk supported languages), please do! Open a pull request or send them to me in a GitHub issue and I'll gladly merge them into the project. I'd like to make this project multi-lingual, but I can't do it without your help.
Every once in a while when downloading a new model on Windows, it will throw an "Access Denied" exception after attempting to rename the extracted file. This isn't really a permissions issue, but rather a concurrency one. I've found that closing any app or Explorer window that might be related to Chapterize-Audiobooks usually fixes this problem. This seems to be a somewhat common issue with Python on Windows when renaming/deleting/moving files.