Pact - Python Audio Clip Tool

Summary: this lets you quickly extract audio clips from .mp3 files (from podcasts, etc) with a transcription, and export it to Anki using Ankiconnect.

mp3 playing and hacking is done using VLC and ffmpeg.

The transcription is done using Vosk.

If you also have a transcription file (in .txt format), it's searched to find the best possible match using the Vosk transcription as a starting point. This is helpful because sometimes the Vosk transcription is a bit off.

NOTE: I've only run this (and tested it) on Mac ... I'm not sure how it will behave on other OS's.

A quick demo

Installation

Installation has 2 parts: parts needed for Pact itself, and then parts needed for your particular use.

Part 1:

• Get the code
• Install VLC
• Install ffmpeg
• Install python requirements
• Test it out

Part 2:

• Install a Vosk model
• AnkiConnect
• Anki notes
• Set up your config.ini

Install 1: Pact

Get the code

# Clone using Github ...
cd /some/directory
git clone [email protected]:jzohrab/pact.git
cd pact

# ... or get it in a zipfile
cd /some/directory
mkdir pact  # or whatever
# get file from https://github.com/jzohrab/pact/archive/refs/heads/main.zip ...
# ... and unzip in pact directory

VLC

Install VLC on your system.

Mac with homebrew:

brew install --cask vlc

Other OSs: ?

ffmpeg

Install ffmpeg on your system.

For Mac with homebrew:

brew install ffmpeg

Other OSs: ?

Install python requirements

python3 -m venv .venv
source .venv/bin/activate

# Install dependencies (some from other git repos, see requirements.txt)
.venv/bin/pip3 install -r requirements.txt

deactivate

or, on systems with bash (Mac etc):

./install-deps.sh

Test it out

At this point, you should be able to run pact:

./pact.sh

The pact app should open. Then you can open an mp3 on your system, or the file "pact/assets/pact.mp3", and click "Play". If you can hear audio, the basics are in place.

Install 2: Your stuff

Vosk model

Vosk uses a "language model" for offline audio transcription.

• Download the appropriate model from https://alphacephei.com/vosk/models

• Unzip it into a subdirectory under the main directory (a sibling to the pact directory).

For example, my directory structure looks like this:

LICENSE
README.md  # (this file)
...
- /model
    - /spanish
        Gr.fst
        HCLr.fst
        README   # (the Vosk model readme)
        disambig_tid.int
        ...
+ /pact
...

Note that some models are very large, and could be slow on your system. For example, the German model vosk-model-de-0.21 is about 2 gigs. When I used it on a brief German clip, the transcription was accurate, but the loading was so slow as to be almost unusable. With the much smaller vosk-model-small-de-0.15 (45 MB), the Vosk transcription was still accurate enough to find the correct match in the accompanying transcription.

NB: Pact currently reloads the entire Vosk model on every transcription; perhaps this could be convered to a one-time load. This would only be useful for very large models.

Ankiconnect

This is an add-on for Anki. See AnkiConnect add-on on AnkiWeb.

See https://github.com/FooSoft/anki-connect#installation for possible additional setup required.

Anki notes

You'll need a Note Type with separate fields for the audio file and the transcription. The names don't matter, because you specify that in the config.ini file.

If you're going to leave the transcription blank -- e.g., if you delete the transcription suggestions that pact provides using Vosk -- the Transcription field can't be the first field in the Note type, because Anki doesn't allow empty first fields.

See the wiki for an Anki card layout idea.

config.ini

Edit config.ini to match your setup.

Plugins

Lookup plugins

Note: I've only used Pact for Spanish, but there are monolingual lookup modules for French and German as well, in ./pact/plugins/lookup/`. One with Python coding experience should be able to implement other dictionary web lookups.

While I haven't needed it myself, Pact should allow for simple plugins for lookups. See ./pact/plugins/lookup/README.md.

If you have a plugin that has the correct function available, you should be able to place it in ./myplugins or in the plugins folder, and then specify the module name in config.ini.

Only use plugins you trust, or that you write yourself ... since they execute arbitrary code, they might be malicious etc etc. Unfortunately it's not easy to sandbox them (so that they can only do "safe" things).

Transcription plugins

Currently, Pact only has a transcription plugin for Vosk, but others should be possible. See ./pact/plugins/transcription/README.md.

Usage

See the pact wiki.

Dev notes

Dependencies

Adding/freezing:

pip3 install X
pip3 freeze > requirements.txt

Dev config

Use PACTCONFIG env variable to use a different config file (e.g., for different Anki export deck, etc.):

PACTCONFIG=config-dev.ini ./pact.sh

Extra config settings:

# Add this section
[Dev]

# Load this session file on start:
SessionFile = /path/to/some/file.clips

# If SessionFile is set, load this bookmark
# (indexed from 0):
LoadBookmark = 7

# Load this mp3 on start, if SessionFile not present
LoadFile = /path/to/file.mp3

Unit tests

Are in /test. Run them all with:

python3 -m venv .venv
source .venv/bin/activate
python -m unittest discover -s test
deactivate

To run a single one matching a pattern, use -k:

python -m unittest discover -s test -k save_session

See python -m unittest -h for more flags.

Releases

• Make sure all tests pass
• Update CHANGELOG.md, pact/_version.py
• Commit
• Tag (e.g. git tag v0.2.0 1f0f2a6)
• git push origin <tagname>

Contributing

• fork the repo
• make a branch, make your changes
• ensure the unit tests all still pass. Add tests as needed to increase coverage/stability
• push to your fork
• open a PR

For some ideas of stuff that might need work, see the wiki.