Releases: alumae/kiirkirjutaja
Releases · alumae/kiirkirjutaja
2.0.1
2.0
- Available as Docker image alumae/kiirkirjutaja:2.0 and alumae/kiirkirjutaja:latest
- Now uses end-to-end streaming transducer models trained using Icefall.
- The new ASR model includes punctuation restoration, i.e., punctuation symbols are automatically inserted to the recognized stream of words
- Should provide 30% less ASR errors and much improved punctuation accuracy
- Docker image size reduced to from ~14 GB to 3.2 GB, as we don't need Kaldi and several other things any more
0.3.0
- Fixed a bug in Youtube presenter that caused timings to be sometimes messed up.
- Introduced a new command line argument
--word-output-delay <seconds>
that can be used to set a minimum time between a word
occurring in speech and word appearing in subtitles. Recommended value is smth like 5. This makes word-by-word output
timing more consistent and makes reading more comfortable. Otherwise, words in utterance ends can be displayed all very 'abruptly'.
Available as Docker image alumae/kiirkirjutaja:0.3.0
and alumae/kiirkirjutaja:latest
. We had to migrate to Docker's native image repository since Koodivaramu doesn't allow to host Docker images any more.
0.2.8
0.2.7
0.2.5
- Fixed a memory leak that caused memory usage to slowly increase
- Increased robustness in cases when FAB/Youtube API endpoint is down or not responding
Available as Docker image: koodivaramu.eesti.ee:5050/taltechnlp/kiirkirjutaja:0.2.5 and koodivaramu.eesti.ee:5050/taltechnlp/kiirkirjutaja:latest
0.2.4
0.2.3
Increased robustness of the language ID module: is a non-target (i.e., not Estonian) language is detected at the 3 second mark, another check is done after 5 seconds. This is to filter out some false negatives when the utterance starts with some some hesitation sounds or otherwise non-standard words that sometimes confuse the language ID module.