Europarl-ASR

Europarl-ASR

Europarl-ASR v1.0
2 April 2021
www.mllp.upv.es/europarl-asr

A large English-language speech and text corpus of parliamentary debates for streaming ASR benchmarking, speech data filtering and speech data verbatimization.

Keywords: automatic speech recognition; speech corpus; speech data filtering; speech data verbatimization.

CONTACT: Gonçal V. Garcés Díaz-Munío (gogardia@vrain.upv.es), Joan Albert Silvestre-Cerdà (jsilvestre@vrain.upv.es). Universitat Politècnica de València.

README CONTENTS

• Overview
• Citation
• Get the data
• Additional Europarl-ASR materials
• Corpus structure and contents
• Extended description
• Acknowledgements
• Legal disclaimers
• Licence

OVERVIEW

Europarl-ASR (EN) includes:

Speech data

• 1300 hours of English-language annotated speech data.
• 3 full sets of timed transcriptions: official non-verbatim transcriptions, automatically noise-filtered transcriptions and automatically verbatimized transcriptions.
• 18 hours of speech data with both manually revised verbatim transcriptions and official non-verbatim transcriptions, split in 2 independent validation- evaluation partitions for 2 realistic ASR tasks (with vs. without previous knowledge of the speaker).

Text data

• 70 million tokens of English-language text data.

Pretrained language models

• The Europarl-ASR English-language n-gram language model and vocabulary.

This data comprises most of the European Parliament's English-language debate recordings, transcriptions and translations available from the Parliament's website from 1996 to 2020. Additionally, to increase text data up to 170M tokens, Europarl-ASR also includes tools to add all English-language text from the DCEP Digital Corpus of the European Parliament.

CITATION

Garcés Díaz-Munío, Gonçal V.; Silvestre-Cerdà, Joan Albert; Jorge, Javier; Giménez, Adrià; Iranzo-Sánchez, Javier; Baquero-Arnal, Pau; Roselló, Nahuel; Pérez-González-de-Martos, Alejandro; Civera, Jorge; Sanchis, Albert; Juan, Alfons. "Europarl-ASR: A Large Corpus of Parliamentary Debates for Streaming ASR Benchmarking and Speech Data Filtering/Verbatimization". In Proc. Interspeech 2021, Brno (Czech Republic), 2021 (in press).

@inproceedings{europarlasr2021,
title = {Europarl-ASR: A Large Corpus of Parliamentary Debates for Streaming ASR Benchmarking and Speech Data Filtering/Verbatimization},
author = {Garcés Díaz-Munío, Gonçal V. and Silvestre-Cerdà, Joan Albert and Javier Jorge and Adrià Giménez and Javier Iranzo-Sánchez and Pau Baquero-Arnal and Nahuel Roselló and Alejandro Pérez-González-de-Martos and Jorge Civera and Albert Sanchis and Alfons Juan},
booktitle = {Proc. Interspeech 2021},
address = {Brno (Czech Republic)},
year = {2021},
keywords = {Automatic Speech Recognition, speech corpus, speech data filtering, speech data verbatimization},
abstract = {[EN] We introduce Europarl-ASR, a large speech and text corpus of parliamentary debates including 1300 hours of transcribed speeches and 70 million tokens of text in English extracted from European Parliament sessions. The training set is labelled with the Parliament’s non-fully-verbatim official transcripts, time-aligned. As verbatimness is critical for acoustic model training, we also provide automatically noise-filtered and automatically verbatimized transcripts of all speeches based on speech data filtering and verbatimization techniques. Additionally, 18 hours of transcribed speeches were manually verbatimized to build reliable speaker-dependent and speaker-independent development/test sets for streaming ASR benchmarking. The availability of manual non-verbatim and verbatim transcripts for dev/test speeches makes this corpus useful for the assessment of automatic filtering and verbatimization techniques. This paper describes the corpus and its creation, and provides off-line and streaming ASR baselines for both the speaker-dependent and speaker-independent tasks using the three training transcription sets. The corpus is publicly released under an open licence.
[CA] "Europarl-ASR: Un extens corpus parlamentari de referència per a reconeixement de la parla i filtratge/literalització de transcripcions": Presentem Europarl-ASR, un extens corpus de veu i text de debats parlamentaris amb 1300 hores d'intervencions transcrites i 70 milions de paraules de text en anglés extrets de sessions del Parlament Europeu. Les transcripcions oficials del Parlament Europeu, no literals, s'han sincronitzat per a tot el conjunt d'entrenament. Com que l'entrenament de models acústics requereix transcripcions com més literals millor, també s'han inclòs transcripcions filtrades i transcripcions literalitzades de totes les intervencions, basades en tècniques de filtratge i literalització automàtics. A més, s'han inclòs 18 hores de transcripcions literals revisades manualment per definir dos conjunts de validació i avaluació de referència per a reconeixement automàtic de la parla en temps real, amb oradors coneguts i amb oradors desconeguts. Pel fet de disposar de transcripcions literals i no literals, aquest corpus és també ideal per a l'anàlisi de tècniques de filtratge i de literalització. En aquest article, es descriu la creació del corpus i es proporcionen mesures de referència de reconeixement automàtic de la parla en temps real i en diferit, amb oradors coneguts i amb oradors desconeguts, usant els tres conjunts de transcripcions d'entrenament. El corpus es fa públic amb una llicència oberta.}
}

GET THE DATA

Download the full Europarl-ASR speech and text corpus from:

https://www.mllp.upv.es/europarl-asr/Europarl-ASR_v1.0.tar.gz

ADDITIONAL Europarl-ASR MATERIALS

In addition to the speech and text data included in the main release and described in this document, we are making available for download the following materials to facilitate the reproducibility of our experiments:

• The pretrained Europarl-ASR English-language n-gram language model, together with its vocabulary file:
  

https://www.mllp.upv.es/europarl-asr/Europarl-ASR_v1.0_ngram_lm_and_vocab.tar.gz

• The Europarl-ASR English-language verbatim transcription guidelines, which were applied to produce the manually revised verbatim transcriptions for the dev and test sets:
  

https://www.mllp.upv.es/europarl-asr/Europarl-ASR_transcription_guidelines.pdf

CORPUS STRUCTURE AND CONTENTS

Total size: 20 GB

The data is organized in 3 main directories: "train" (training data), "dev" (validation data) and "test" (evaluation data). Each directory contains the subdirectories "original_audio" and "text", the first one containing speech data with annotations (for acoustic modelling), the second one containing text data (for language modelling).

Here we can see more completely the corpus structure, with additional subdirectories:

  Europarl-ASR
  └── en
      ├── dev
      │   ├── original_audio
      │   │   ├── spk-dep
      │   │   │   ├── lists
      │   │   │   ├── metadata
      │   │   │   ├── refs
      │   │   │   └── speeches
      │   │   └── spk-indep
      │   │       ├── lists
      │   │       ├── metadata
      │   │       ├── refs
      │   │       └── speeches
      │   └── text
      │       ├── prepro
      │       └── raw
      ├── test
      │   ├── original_audio
      │   │   ├── spk-dep
      │   │   │   ├── lists
      │   │   │   ├── metadata
      │   │   │   ├── refs
      │   │   │   └── speeches
      │   │   └── spk-indep
      │   │       ├── lists
      │   │       ├── metadata
      │   │       ├── refs
      │   │       └── speeches
      │   └── text
      │       ├── prepro
      │       └── raw
      └── train
          ├── original_audio
          │   ├── lists
          │   ├── metadata
          │   └── speeches
          └── text
              ├── external
              │   ├── prepro
              │   └── scripts
              └── internal
                  ├── prepro
                  └── raw

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

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", "refs" (only in "dev" and "test") and "speeches".

"lists" contains lists of all the speeches in "speeches", as well as lists of speeches per speaker.

"metadata" contains metadata for each speech and for each speaker in the 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

And for each speaker (as reflected in speakers.headers.csv):

    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>/

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>.tr.orig.{dfxp,json,srt,txt}
    [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 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 a txt transcription file, as dfxp or srt 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, as a txt transcription file, as dfxp or srt timed subtitle files, and its json metadata.

Finally, in "refs" (only in "dev" and "test") each file contains every speech in the corresponding dev or test set, that is, the full reference for that set. In each case, we will find 4 files, containing the official non-verbatim reference (*.orig.*) and the manually revised verbatim reference (*.rev.*), as 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)

In the case of "train", they are subdivided in directories "external" and "internal". "internal" contains all the official non-verbatim transcriptions and translations in the train set, together with the selected non-overlapping Europarl v10 transcriptions and translations; "external" contains the files to make use of the external DCEP: Digital Corpus of the European Parliament.

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 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 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 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).

EXTENDED DESCRIPTION

Europarl-ASR (EN) is a large English-language speech and text corpus of parliamentary debates for (streaming) ASR benchmarking and speech data filtering/verbatimization, including 1300 hours of annotated English-language speeches from European Parliament sessions held in the period 1996-2020.

It was compiled and released by the Machine Learning and Language Processing (MLLP) research group of VRAIN Institut Valencià d'Investigació en Intel·ligència Artificial, Universitat Politècnica de València ( www.mllp.upv.es ).

Europarl-ASR (EN) includes:

Speech data

• 1300 hours of English-language annotated speech data (33K speeches, 1K speakers).
• 3 full sets of timed transcriptions: official non-verbatim transcriptions, automatically noise-filtered transcriptions and automatically verbatimized transcriptions.
• 18 hours of speech data with both manually revised verbatim transcriptions and official non-verbatim transcriptions, split in 2 independent validation- evaluation partitions for 2 realistic ASR tasks (with vs. without previous knowledge of the speaker).

Text data

• 70 million tokens of English-language text data.

Language models

• The Europarl-ASR English-language n-gram language model and vocabulary.

This data comprises most of the European Parliament's English-language debate recordings, transcriptions and translations available from the Parliament's website from 1999 to 2020 (recordings being only available from 2008). This is complemented by including all English-language transcriptions and translations from the Europarl v10 text corpus for the period 1996-1999.

Additionally, to increase text data for language modelling up to 170M tokens, Europarl-ASR also includes tools to add all English-language text from the DCEP Digital Corpus of the European Parliament.

Detailed dates of the EP speech and text data gathered:

• English speech: 2008-09-01 to 2020-05-27.
• English transcriptions: 1999-07-20 to 2020-05-27.
• Translations into English: 1999-07-20 to 2012-11-30.
• Europarl v10 (selected to avoid overlapping): 1996-04-15 to 1999-07-19.
• DCEP (does not include any EP reports of proceedings): 2001 to 2012.

ACKNOWLEDGEMENTS

The authors would like to acknowledge:

• The European Parliament, the European Commission and other EU organizations, for making available a wealth of multilingual speech and text data, both in their websites and as ready-made corpora such as the DCEP Digital Corpus of the European Parliament ( https://ec.europa.eu/jrc/en/language-technologies ).

• Philipp Koehn for compiling the Europarl corpus ( 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, MCIU/AEI/FEDER,EU) and FPU scholarships FPU14/03981 and FPU18/04135; the Generalitat Valenciana's research project Classroom Activity Recognition (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

• Speech and text data from the European Parliament website (audio, official transcriptions and translations) were sourced from https://www.europarl.europa.eu/plenary/en/debates-video.html

• Text data from the DCEP Digital Corpus of the European Parliament are the exclusive property of the European Parliament. These data were sourced from https://ec.europa.eu/jrc/en/language-technologies/dcep (date of the latest update: 11 March 2015).

LICENCE

• Speech and text data from the European Parliament website (audio, official transcriptions and translations) are the exclusive property of the European Union represented by the European Parliament. These data are reused here under the conditions stated in the European Parliament website's Legal notice ( https://www.europarl.europa.eu/legal-notice ).

• Text data from the DCEP Digital Corpus of the European Parliament are the exclusive property of the European Parliament. The European Parliament retains ownership of the data. These data are reused here under the usage conditions of the DCEP Digital Corpus of the European Parliament ( https://ec.europa.eu/jrc/en/language-technologies/dcep#Usage%20Conditions ).

• Text data from the Europarl v10 corpus are reused here under the Europarl corpus terms of use ( https://www.statmt.org/europarl/ ).

• Europarl-ASR data and code not covered by the previously mentioned licences © 2021 by Pau Baquero-Arnal, Jorge Civera, Gonçal V. Garcés Dı́az-Munı́o, Adrià Giménez, Javier Iranzo-Sánchez, Javier Jorge, Alfons Juan, Alejandro Pérez-González-de-Martos, Nahuel Roselló, Albert Sanchis and Joan Albert 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 LICENSE file for the full licence texts.