Code for processing the CHiME-5 signal to:
- i/ Fix frame dropping misalignment by 0 insertion
- ii/ Compensation for clock drift and start-time mis-estimation using sox
- Python 3.5+
- Sox v14.4.2 (We confirmed that it produces the different result with v14.4.1. Please make sure to use v14.4.2, which is the latest version)
./run_all.sh
The op level script which will
- correct the audio by calling
correct_audio.sh
once for each of the 20 sessions. Currently usingqsub
. - correct the transcripts called
correct_transcript_for_clock_drift.py
correct_audio.sh
processes a single CHiME session in two steps calling
- correct_signals_for_frame_drops.py
- correct_signals_for_clock_drift.py Should not take more than 30-40 mins per session.
Note, the tools are making the corrections according to data stored in chime6_audio_edits.json
. These data has been derived using tools operating on the audio from the AVI video files that have not been released.
- Check paths.sh and set the paths for CHiME5, CHiME6 and sox
- You'll need a python3 environment with modules listed in requirements.txt
audio_md5sums.txt
- md5sums for all the processed audio files, you can check it with the following command:
. ./paths.sh
cd $CHiME6_ROOT/../
md5sum -c audio_md5sums.txt
If your processing is successfully done, you'll get the following results:
CHiME6/audio/train/S08_U02.CH3.wav: OK
CHiME6/audio/train/S08_U03.CH3.wav: OK
CHiME6/audio/dev/S09_U01.CH3.wav: OK
CHiME6/audio/train/S19_U04.CH1.wav: OK
CHiME6/audio/train/S05_U01.CH1.wav: OK
:
- S01_U03 - missing because video software failed to record any audio. (unrecoverable)
- S05_U04 - missing, video software was temporarily stopped during recording.