|
@@ -1,7 +1,7 @@
|
|
|
-Europarl-ASR
|
|
|
-v1.0
|
|
|
-2 April 2021
|
|
|
-www.mllp.upv.es/europarl-asr/
|
|
|
+# Europarl-ASR
|
|
|
+v1.0<br />
|
|
|
+2 April 2021<br />
|
|
|
+www.mllp.upv.es/europarl-asr
|
|
|
|
|
|
A large English-language speech and text corpus of parliamentary debates for
|
|
|
streaming ASR benchmarking and speech data filtering/verbatimization.
|
|
@@ -18,21 +18,21 @@ Universitat Politècnica de València.
|
|
|
README CONTENTS
|
|
|
---------------
|
|
|
|
|
|
-- Overview
|
|
|
-- Corpus structure and contents
|
|
|
-- Additional Europarl-ASR materials
|
|
|
-- Extended description
|
|
|
-- Acknowledgements
|
|
|
-- Legal disclaimers
|
|
|
-- Licence
|
|
|
+- [Overview](#overview)
|
|
|
+- [Corpus structure and contents](#contents)
|
|
|
+- [Additional Europarl-ASR materials](#additional)
|
|
|
+- [Extended description](#description)
|
|
|
+- [Acknowledgements](#ack)
|
|
|
+- [Legal disclaimers](#legal)
|
|
|
+- [Licence](#licence)
|
|
|
|
|
|
|
|
|
-OVERVIEW
|
|
|
+<a id="overview"></a>OVERVIEW
|
|
|
--------
|
|
|
|
|
|
Europarl-ASR (EN) includes:
|
|
|
|
|
|
-*Speech data:
|
|
|
+#### Speech data
|
|
|
|
|
|
- 1.3K hours of English-language annotated speech data.
|
|
|
- 18 hours of speech data with both manually revised verbatim transcriptions
|
|
@@ -43,11 +43,11 @@ Europarl-ASR (EN) includes:
|
|
|
(training partition): official non-verbatim transcriptions, automatically
|
|
|
noise-filtered transcriptions and automatically verbatimized transcriptions.
|
|
|
|
|
|
-*Text data:
|
|
|
+#### Text data
|
|
|
|
|
|
- 70M tokens of English-language text data.
|
|
|
|
|
|
-*Pretrained language models:
|
|
|
+#### Pretrained language models
|
|
|
|
|
|
- The Europarl-ASR English-language n-gram language model and vocabulary.
|
|
|
|
|
@@ -58,7 +58,7 @@ tokens, Europarl-ASR also includes tools to add all English-language text from
|
|
|
the DCEP Digital Corpus of the European Parliament.
|
|
|
|
|
|
|
|
|
-CORPUS STRUCTURE AND CONTENTS
|
|
|
+<a id="contents"></a>CORPUS STRUCTURE AND CONTENTS
|
|
|
-----------------------------
|
|
|
|
|
|
Total size: 20 GB
|
|
@@ -72,6 +72,7 @@ data (for language modelling).
|
|
|
Here we can see more completely the corpus structure, with additional
|
|
|
subdirectories:
|
|
|
|
|
|
+```
|
|
|
Europarl-ASR
|
|
|
└── en
|
|
|
├── dev
|
|
@@ -116,18 +117,19 @@ subdirectories:
|
|
|
└── internal
|
|
|
├── prepro
|
|
|
└── raw
|
|
|
+```
|
|
|
|
|
|
-*Speech data ("original_audio" directories):
|
|
|
+#### Speech data ("original_audio" directories)
|
|
|
|
|
|
In the cases of "dev" and "test", they are subdivided in directories "spk-dep"
|
|
|
and "spk-indep". Thus, for speech data, we have 2 train-dev-test partitions
|
|
|
for 2 different ASR tasks, as follows:
|
|
|
|
|
|
- 1) ASR with known speakers (MEP):
|
|
|
- train ; dev/original_audio/spk-dep ; test/original_audio/spk-dep
|
|
|
-
|
|
|
- 2) ASR with unknown speakers (Guest):
|
|
|
- train ; dev/original_audio/spk-indep ; test/original_audio/spk-indep
|
|
|
+1. ASR with known speakers (MEP):<br />
|
|
|
+ train ; dev/original_audio/spk-dep ; test/original_audio/spk-dep
|
|
|
+
|
|
|
+1. ASR with unknown speakers (Guest):<br />
|
|
|
+ train ; dev/original_audio/spk-indep ; test/original_audio/spk-indep
|
|
|
|
|
|
Each of these partition directories contains 3 to 4 subdirectories (depending
|
|
|
on whether it is the train set or the dev/test sets): "lists", "metadata",
|
|
@@ -140,40 +142,40 @@ speeches per speaker.
|
|
|
corresponding set (as csv and json files). For each speech we will find these
|
|
|
metadata (as reflected in speeches.headers.csv):
|
|
|
|
|
|
- term;session_date;speech_id;speaker_type;speaker_id;raw_dur;
|
|
|
- aligned-speech_dur;filtered-speech_dur;cer;ar;path;agenda_item_title
|
|
|
+ term;session_date;speech_id;speaker_type;speaker_id;raw_dur;<br />
|
|
|
+ aligned-speech_dur;filtered-speech_dur;cer;ar;path;agenda_item_title
|
|
|
|
|
|
And for each speaker (as reflected in speakers.headers.csv):
|
|
|
|
|
|
- type;id;name;gender;url
|
|
|
+ type;id;name;gender;url
|
|
|
|
|
|
"speeches" contains a subdirectory for each speech in the corresponding set,
|
|
|
according to this subdirectory structure:
|
|
|
|
|
|
- speeches/<term>/<session_date>/<speech_id>/
|
|
|
+ `speeches/<term>/<session_date>/<speech_id>/`
|
|
|
|
|
|
For each speech, we will find some of the following files (depending on
|
|
|
whether it is in the train set or in the dev/test sets):
|
|
|
|
|
|
- ep-asr.en.orig.<term>.<session_date>.<speech_id>.m4a
|
|
|
- [In all sets] Audio of the speech.
|
|
|
+ `ep-asr.en.orig.<term>.<session_date>.<speech_id>.m4a`<br />
|
|
|
+ [In all sets] Audio of the speech.
|
|
|
|
|
|
- ep-asr.en.orig.<term>.<session_date>.<speech_id>.tr.orig.{dfxp,json,srt,txt}
|
|
|
- [In all sets] Official non-verbatim transcription of the speech, as a txt
|
|
|
+ `ep-asr.en.orig.<term>.<session_date>.<speech_id>.tr.orig.{dfxp,json,srt,txt}`<br />
|
|
|
+ [In all sets] Official non-verbatim transcription of the speech, as a txt
|
|
|
raw transcription file, as dfxp or srt force-aligned timed subtitle files,
|
|
|
and its json metadata.
|
|
|
|
|
|
- ep-asr.en.orig.<term>.<session_date>.<speech_id>.tr.filt.{dfxp,json,srt}
|
|
|
- [In train set] Automatically filtered transcription of the speech, as dfxp
|
|
|
+ `ep-asr.en.orig.<term>.<session_date>.<speech_id>.tr.filt.{dfxp,json,srt}`<br />
|
|
|
+ [In train set] Automatically filtered transcription of the speech, as dfxp
|
|
|
or srt force-aligned timed subtitle files, and its json metadata.
|
|
|
|
|
|
- ep-asr.en.orig.<term>.<session_date>.<speech_id>.tr.verb.{dfxp,json,srt,txt}
|
|
|
- [In train set] Automatically verbatimized transcription of the speech, as
|
|
|
+ `ep-asr.en.orig.<term>.<session_date>.<speech_id>.tr.verb.{dfxp,json,srt,txt}`<br />
|
|
|
+ [In train set] Automatically verbatimized transcription of the speech, as
|
|
|
a txt transcription file, as dfxp or srt force-aligned timed subtitle files,
|
|
|
and its json metadata.
|
|
|
|
|
|
- ep-asr.en.orig.<term>.<session_date>.<speech_id>.tr.rev.{dfxp,json,srt,txt}
|
|
|
- [In dev/test sets] Manually revised verbatim transcription of the speech,
|
|
|
+ `ep-asr.en.orig.<term>.<session_date>.<speech_id>.tr.rev.{dfxp,json,srt,txt}`<br />
|
|
|
+ [In dev/test sets] Manually revised verbatim transcription of the speech,
|
|
|
as a txt transcription file, as dfxp or srt force-aligned timed subtitle
|
|
|
files, and its json metadata.
|
|
|
|
|
@@ -185,7 +187,7 @@ transcriptions (*.ref) and as segment time marked files (*.stm). In all 4
|
|
|
cases, the text is presented preprocessed for evaluation (tokenized,
|
|
|
lowercased, punctuation removed...).
|
|
|
|
|
|
-*Text data ("text" directories):
|
|
|
+#### Text data ("text" directories)
|
|
|
|
|
|
In the case of "train", they are subdivided in directories "external" and
|
|
|
"internal". "internal" contains all the official non-verbatim transcriptions
|
|
@@ -197,24 +199,24 @@ Each "text" directory contains 2 subdirectories: "raw" (except in
|
|
|
"train/external"), "prepro" (in all sets), or "scripts" (only in
|
|
|
"train/external").
|
|
|
|
|
|
- "raw" contains the raw text data for the corresponding set (*.txt.gz), and
|
|
|
+ "raw" contains the raw text data for the corresponding set (*.txt.gz), and
|
|
|
its metadata (*.csv). In the cases of "dev" and "test", both the official
|
|
|
non-verbatim transcriptions (*.orig.*) and the manually revised verbatim
|
|
|
transcriptions (*.rev.*) are included.
|
|
|
|
|
|
- "prepro" contains the text data for the corresponding set, preprocessed for
|
|
|
+ "prepro" contains the text data for the corresponding set, preprocessed for
|
|
|
training or evaluation (tokenized, lowercased, punctuation removed...). This
|
|
|
data is released to facilitate the reproducibility of our experiments.
|
|
|
|
|
|
- Finally, "scripts" (only in "train/text/external") contains the script
|
|
|
+ Finally, "scripts" (only in "train/text/external") contains the script
|
|
|
get_DCEP.sh, which can be used to download the DCEP corpus from its original
|
|
|
website and save it in compressed plain text (.txt.gz).
|
|
|
|
|
|
|
|
|
-ADDITIONAL Europarl-ASR MATERIALS
|
|
|
+<a id="additional-materials"></a>ADDITIONAL Europarl-ASR MATERIALS
|
|
|
---------------------------------
|
|
|
|
|
|
-https://www.mllp.upv.es/europarl-asr/Europarl-ASR_v1.0_ngram_lm_and_vocab.tar.gz
|
|
|
+https://www.mllp.upv.es/europarl-asr/Europarl-ASR_v1.0_ngram_lm_and_vocab.tar.gz<br />
|
|
|
https://www.mllp.upv.es/europarl-asr/Europarl-ASR_transcription_guidelines.pdf
|
|
|
|
|
|
In addition to the speech and text data included in the main release and
|
|
@@ -229,7 +231,7 @@ materials to facilitate the reproducibility of our experiments:
|
|
|
dev and test sets.
|
|
|
|
|
|
|
|
|
-EXTENDED DESCRIPTION
|
|
|
+<a id="description"></a>EXTENDED DESCRIPTION
|
|
|
--------------------
|
|
|
|
|
|
Europarl-ASR (EN) is a large English-language speech and text corpus of
|
|
@@ -244,7 +246,7 @@ Intel·ligència Artificial, Universitat Politècnica de València
|
|
|
|
|
|
Europarl-ASR (EN) includes:
|
|
|
|
|
|
-*Speech data:
|
|
|
+#### Speech data
|
|
|
|
|
|
- 1.3K hours of English-language annotated speech data (33K speeches, 1K
|
|
|
speakers).
|
|
@@ -256,11 +258,11 @@ Europarl-ASR (EN) includes:
|
|
|
(training partition): official non-verbatim transcriptions, automatically
|
|
|
noise-filtered transcriptions and automatically verbatimized transcriptions.
|
|
|
|
|
|
-*Text data:
|
|
|
+#### Text data
|
|
|
|
|
|
- 70M tokens of English-language text data.
|
|
|
|
|
|
-*Language models:
|
|
|
+#### Language models
|
|
|
|
|
|
- The Europarl-ASR English-language n-gram language model and vocabulary.
|
|
|
|
|
@@ -283,7 +285,7 @@ Detailed dates of the EP speech and text data gathered:
|
|
|
- DCEP (does not include any EP reports of proceedings): 2001 to 2012.
|
|
|
|
|
|
|
|
|
-ACKNOWLEDGEMENTS
|
|
|
+<a id="ack"></a>ACKNOWLEDGEMENTS
|
|
|
---------------
|
|
|
|
|
|
The authors would like to acknowledge:
|
|
@@ -298,15 +300,15 @@ The authors would like to acknowledge:
|
|
|
( https://www.statmt.org/europarl/ ).
|
|
|
|
|
|
This work has received funding from the EU's H2020 research and innovation
|
|
|
-programme under grant agreements 761758 (X5gon) and 952215 (TAILOR); the
|
|
|
-Government of Spain's research project Multisub (RTI2018-094879-B-I00,
|
|
|
+programme under grant agreements 761758 ([X5gon](https://cordis.europa.eu/project/id/761758)) and 952215 ([TAILOR](https://cordis.europa.eu/project/id/952215)); the
|
|
|
+Government of Spain's research project [Multisub](https://www.mllp.upv.es/projects/multisub/) (RTI2018-094879-B-I00,
|
|
|
MCIU/AEI/FEDER,EU) and FPU scholarships FPU14/03981 and FPU18/04135; the
|
|
|
-Generalitat Valenciana's research project Classroom Activity Recognition
|
|
|
+Generalitat Valenciana's research project [Classroom Activity Recognition](https://aplicat.upv.es/exploraupv/ficha-proyecto/proyecto/20190714)
|
|
|
(PROMETEO/2019/111) and predoctoral research scholarship ACIF/2017/055; and
|
|
|
the Universitat Politècnica de València's PAID-01-17 R&D support programme.
|
|
|
|
|
|
|
|
|
-LEGAL DISCLAIMERS
|
|
|
+<a id="legal"></a>LEGAL DISCLAIMERS
|
|
|
-----------------
|
|
|
|
|
|
- Speech and text data from the European Parliament website (audio, official
|
|
@@ -319,7 +321,7 @@ LEGAL DISCLAIMERS
|
|
|
update: 11 March 2015).
|
|
|
|
|
|
|
|
|
-LICENCE
|
|
|
+<a id="licence"></a>LICENCE
|
|
|
-------
|
|
|
|
|
|
- Speech and text data from the European Parliament website (audio, official
|
|
@@ -344,4 +346,4 @@ LICENCE
|
|
|
Silvestre-Cerdà are licenced under CC BY 4.0. To view a copy of this
|
|
|
licence, visit http://creativecommons.org/licenses/by/4.0/
|
|
|
|
|
|
-See the file LICENSE for the full licence texts.
|
|
|
+See the [LICENSE](https://mllp.upv.es/git-pub/ggarces/Europarl-ASR/src/master/LICENSE) file for the full licence texts.
|