Evaluation and development data sets for speech translation for meetings

Authors: Pinnis, Mārcis ; Pole, Megija ; Kapočiūtė-Dzikienė, Jurgita ; Nicmanis, Dāvis ; Salimbajevs, Askars ; Skadiņš, Raivis ; Miķelsons, Mārtiņš ; Lizanders, Kristaps ; Bērziņš, Aivars ; Vasiļevskis, Artūrs ; Rozis, Roberts; Kornikaite, Nida

License: Creative Commons - Attribution-ShareAlike 4.0 International (CC BY-SA 4.0)

Description of data formats

The repository consists of the following:
1) "Speech annotation guidelines - public.docx" - Guidelines that the annotators used to annotate the evaluation and development data sets for speech translation for meetings
2) "lt-en.zip" - the Lithuanian->English evaluation and development data
3) "lv-en-dev.zip" - the Latvian->English development data
4) "lv-en-eval.zip" - the Latvian->English evaluation data
5) "en-lv-dev.zip" - the English->Latvian development data
6) "en-lv-eval.zip" - the English->Latvian evaluation data

The directory structure of the ZIP archives is as follows:
1) "lt-en.zip":
	dev
		*.antx
		*.wav
	eval
		*.antx
		*.wav
	lt_data_dev.tsv
	lt_data_eval.tsv
2) "lv-en-dev.zip":
	dev
		*.antx
		*.wav
	inter-annotator-agreement
		a1
			*.antx
		a2
			*.antx
	lv_data_dev.tsv
3) "lv-en-eval.zip":
	eval
		*.antx
		*.wav
	lv_data_eval.tsv
4) "en-lv-dev.zip":
	dev
		*.antx
		*.wav
	en_data_dev.tsv
5) "en-lv-eval.zip":
	eval
		*.antx
		*.wav
	en_data_eval.tsv

Each audio file (in the WAV data format) has been annotated by one annotator. The annotation was performed using Annotation Pro (http://annotationpro.org/). Annotations for each annotated WAV file were saved in an Annotation Pro XML Annotation File (ANTX file). The file names for each *antx and *wav file pair are equal.

Then, for each translation direction, we have also extracted all annotations and stored them in tab-separated values (TSV) files (i.e., the [lt|en|lv]_data[dev|eval].tsv files). Each line in the TSV files corresponds to one annotated segment/sentence. Each segment/sentence consists of the following tab-separated values:
1) Name of the WAV/ANTX file (without file extension).
2) ID of the annotator (annotator IDs are reset for each translation direction).
3) Start time (in seconds) of the segment (in the particular WAV file).
4) End time (in seconds) of the segment (in the particular WAV file).
5) ID of the speaker (speaker IDs are reset for each WAV file; i.e., they always start from S1, S2, ...).
6) Orthographic transcription.
7) Normalised transcription (with spoken language slang and mispronounced words normalised into written language).
8) Trimmed transcription (with verbal noise, partially spoken words, phenomena non-existing in written language deleted).
9) Reordered transcription (such that the word order would better adhere to the correct word order of the source language).
10) Translation (either into English (for Latvian and Lithuanian source data) or Latvian (for English source data)).