scripts/README.md

Scripts

make_clips.py: cut the clips from source videos

The dataset contains no audio. make_clips.py cuts the clips of the dataset from source videos that you have downloaded yourself, with the same processing that produced the dataset:

1. decode the audio track to 16 kHz mono 16-bit PCM with ffmpeg (ffmpeg -vn -ac 1 -ar 16000 -c:a pcm_s16le);
2. take the samples from int(start * 16000) to int(end * 16000);
3. write them as FLAC to <out-dir>/<audio_input_path>.

The script does not download videos. The source video of a row is https://www.youtube.com/watch?v=<video_id>.

Requirements

Python 3.9 or later with pyarrow and soundfile, and the ffmpeg program:

pip install pyarrow soundfile

Usage

Get the dataset rows, for example the Dutch rows of the asr split (files data/asr/nl/):

huggingface-cli download meetween/mumospee_v3 --repo-type dataset \
    --include "data/asr/nl/*" --local-dir mumospee

Cut the clips of one video:

python make_clips.py --rows mumospee/data/asr/nl \
    --video downloads/-H9G_NHn4Xc.mp4 --out-dir clips

or of all videos in a folder:

python make_clips.py --rows mumospee/data/asr/nl --videos-dir downloads --out-dir clips

The video ID is read from the file name: <video_id>.<ext>, <video_id>_<anything>.<ext> or <anything>[<video_id>].<ext>; otherwise pass --video-id. --rows also accepts a CSV or JSON Lines file with the columns video_id, clip_id, start, end and audio_input_path.

Output

clips/
├── audio/<lang>/<video_id>/<clip_id>.flac   one file per clip, at the row's audio_input_path
└── make_clips_report.csv                    video_id, clip_id, status (written / skipped / failed), note

The audio of a row is clips/<audio_input_path>. Clips that already exist are skipped, so an interrupted run can be repeated; --overwrite rewrites them.

Notes

• The clips match the clips the dataset was built from when the downloaded audio is the same as the original. In a test with the original source files, 142 of 144 clips were identical sample by sample, and the other 2 differed by at most one unit in single samples (a rounding difference between ffmpeg versions).
• A video downloaded at a different time or in another format can differ slightly from the original, and a video may no longer be available.
• If the downloaded audio is shorter than a clip's end, the clip is cut at the end of the audio, and the report notes this.