Datasets:
license: cc-by-sa-4.0
task_categories:
- automatic-speech-recognition
language:
- et
Transcribed ERR Video News Dataset
This dataset contains transcriptions of video news stories from Estonian National Brroacasting (https://www.err.ee/). There are around 40K stories with a total duration of around 4000 hours. Transcriptions are generated automatically using speech recognition (gemini-3-flash-preview). Contextual biasing was used to improve ASR quality, using the textual news story about the same topic. The WER of the transcriptions is around 5% on the average.
The dataset has been heavily filtered to contain only those stories where the corresponding video contains mostly Estonian speech. That is, news pieces that contain a lot of non-Estonian speech and/or music have been removed.
For each story, the dataset provides heading, leadin text and the main body of the textual news story, as well as the transcript of the video news story. In addition, there is also the "subtitles" field, that contains the same text as in the "transcript" field, but segmented into VTT-formatted subtitles, where each subtitle block represents one sentence. Sentence/subtitle start and end times are obtained using forced alignment.
In order to avoid copyright problems, audio/video data is not included in the dataset. However, links to web pages where the original video can be scraped is provided for each story. Contact tanel.alumae@taltech.ee if you need help with downloading the audio.