This repository has scripts to extract data from various sources for ASR and speaker diarization.
It was created to normalize the text by RÚV but it should be easy to use it to normalize other text as well, given that each segment has a segmentID
Extract subtitle and video files from the RÚV api
Cleaning subtitle files and converting video to audio only
Given a subtitle file, segments and text files for ASR can be created. The script, create_segments_and_text.py assumes that the subtitle file is in a directory named after the show. The segments and text files will appear in in the existing local data directory.
Currently this repository works with Python 3.
888 is the name for the RÚV teletext subtitles. It is what needs to be pressed to make the subtitles appear on a TV screen. Expanded 888 are subtitles which have been text normalized to remove capitalizations, numbers(digits), and punctuation.
Template versions of the json files have been uploaded to the json folder. All the values in root_urls.json need to be filled out to extract anything from the api.
After the subtitle files are cleaned, they can be compared to automatic speech recognition(ASR) output using compare_hypothesis_and_expanded_888.sh
To use, compare_hypothesis_and_expanded_888.sh, you need to set up kaldi. Then, within path.sh, change /data/kaldi to the path of your version of kaldi.
Both the ASR transcript and the expanded 888 need to have reference ids at the beginning of every line.
python3 extract_from_ruv_api.py --url-file json/root_urls.json --shows-file json/all_shows.json
python3 create_segments_and_text.py --subtitle-file /absolute/path/for/webvtt/file
./compare_hypothesis_and_expanded_888.sh -h
./compare_hypothesis_and_expanded_888.sh data/ASR/transcript.txt data/ref/expanded_no_punct.txt
ELAN is a professional tool for creating annotations for audio or video. Extract data from ELAN files. ELAN is a very populator annotation tool for conversational annotators. It works on top of PRAAT.
files contains the speakerIDS can use this to create reco2num_spk
contain the segments/time_slots, speakerIDs, and transcriptions along with annotations
contains the audio of the corresponding conversation
Developer
- Judy Fong - [email protected]
This is part of the Language Technology Program by The Icelandic Government through Almannaromur.