https://github.com/meyskens/open-emma
Open Library of EMMA recordings
https://github.com/meyskens/open-emma
Last synced: about 1 month ago
JSON representation
Open Library of EMMA recordings
- Host: GitHub
- URL: https://github.com/meyskens/open-emma
- Owner: meyskens
- Created: 2020-04-10T18:30:29.000Z (about 5 years ago)
- Default Branch: master
- Last Pushed: 2020-04-24T17:54:48.000Z (about 5 years ago)
- Last Synced: 2025-01-22T05:17:08.657Z (3 months ago)
- Size: 56.5 MB
- Stars: 3
- Watchers: 4
- Forks: 0
- Open Issues: 3
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# Open Source EMMA
EMMA (Electronic Management of Message and Announcement) is the announcement system of NMBS/SNCB and Infrabel (ownership changed over time).
This projet aims to be an open library of audio messages by EMMA, for use by train fans.
## More about EMMA
* [Video Infrabel NL](https://www.youtube.com/watch?v=MFhZfuNl2iM)
* [Video SNCB Inside](https://www.youtube.com/watch?v=qnrqK25dHvM)
## How this works
Since no official recordings are available we rely on recorings inside train stations.
These have a lot of background noise, this we can filter out with the help of a Neural Net.
We use [Looking to Listen](https://github.com/meokz/looking-to-listen), we use a modification of the Google Research to do an audtio only analysis.
Cleaned up recordings are placed in `cleaned/`, these are then cut into pieces with the best recording of each word.
### How to clean up recordings
You will need a sound recording of EMMA in WAV format.
You will also need a copy of https://github.com/meokz/looking-to-listen. Note: this uses a lot of RAM, running it in a cloud instance is adviced.
The neural net has issues processing silence, this can be filtered using ffmpeg:
```
ffmpeg -i IN.wav -af "silenceremove=start_periods=1:start_duration=1:start_threshold=-60dB:detection=peak,aformat=dblp,areverse,silenceremove=start_periods=1:start_duration=1:start_threshold=-60dB:detection=peak,aformat=dblp,areverse" OUT.wav
```
After that you can place it inside `./data/noice/` of the looking-to-listen directory and launch it using
```
docker-compose run network python3 quick_start_audio_only.py /data/model/0f_1sclean_noise.npz /data/noise -ideep
```
The cleaned up can be found in `./data/result/name/clean.wav`.
## Project structure
`cleaned/` contains long form audio samples which are cleaned up by the neural net.
`en/ nl/ fr/ de/` are the languages folders.
In each of these are sub-categories:
* `numbers` for numbers
* `words` general words or sentances to composite messages
* `trains` for train types eg. IC, IR, L, S1, Thalys, Eurostar
* `stations` contains station names in telegegraph notation (see http://users.telenet.be/pk/stations.htm) in case of doubt (Noorderkempen) use the NMBS notation. In case of no code (non-Belgian stations) the name in it's native language will be used.
(structure is work in progress)
for versioning of old recordings we should trust Git and not rely on versioning in file names.
We are currently using 48000Hz 16bit PCM WAV to store the recordings. While not space efficient it is commonly used when programming with audio.