In this newsletter:

Membership Year 2020 Publication Preview
LDC Data and Commercial Technology Development

New Publications:
BOLT English Treebank - Discussion Forum
Polish Speech Database
2016 NIST Speaker Recognition Evaluation Test Set

Membership Year 2020 Publication Preview

The 2020 Membership Year is just around the corner and plans for next year’s publications are in progress. Among the expected releases are:

Abstract Meaning Representation (AMR) Annotation Release 3.0: semantic treebank of over 59,000 English natural language sentences from broadcast conversations, newswire, weblogs and web discussion forums; updates the second version (LDC2017T10) with new annotations.

TAC KBP: English sentiment slot filling, surprise slot filling, nugget detection and coreference, and event argument data in all languages (English, Chinese, and Spanish)

DEFT Chinese ERE: Chinese discussion forum data annotated for entities, relations, and events

LibriVox Spanish: 73 hours of Spanish audiobook read speech and transcripts

IARPA Babel Language Packs (telephone speech and transcripts): languages include Dhuluo, Javanese, and Mongolian

HAVIC Med Training data: web video, metadata, and annotations for developing multimedia systems

RATS Speaker Identification: conversational telephone speech in Levantine Arabic, Pashto, Urdu, Farsi and Dari on degraded audio signals with annotation of speech segments for speaker identification

BOLT: discussion forums, SMS/chat, conversational telephone speech, word-aligned, tagged and co-reference data in all languages (Chinese, Egyptian Arabic, and English)

Check your inbox in the coming weeks for more information about membership renewal. 

LDC data and commercial technology development

For-profit organizations are reminded that an LDC membership is a pre-requisite for obtaining a commercial license to almost all LDC databases. Non-member organizations, including non-member for-profit organizations, cannot use LDC data to develop or test products for commercialization, nor can they use LDC data in any commercial product or for any commercial purpose. LDC data users should consult corpus-specific license agreements for limitations on the use of certain corpora. Visit the Licensing page for further information.

New publications:

(1) BOLT English Treebank - Discussion Forum was developed by LDC and consists of 268,907 tokens of English web discussion forum data with part-of-speech and syntactic structure annotations collected for the DARPA BOLT (Broad Operational Language Translation) program.

Part-of-speech and treebank annotation conformed to Penn Treebank II style, incorporating changes to those guidelines that were developed under the GALE (Global Autonomous Language Exploitation) program. Supplementary guidelines for English treebanks and web text are included with this release.

The source data is English discussion forum web text collected by LDC in 2011 and 2012. A subset of that data -- 702 files representing 268,907 tokens -- was selected for the treebank and annotated for word-level tokenization, part-of-speech and syntactic structure. The unannotated English source data is released as BOLT English Discussion Forums (LDC2017T11).

BOLT English Treebank - Discussion Forum is distributed via web download.

2019 Subscription Members will automatically receive copies of this corpus. 2019 Standard Members may request a copy as part of their 16 free membership corpora. Non-members may license this data for $2000.

*

(2) Polish Speech Database was developed by VoiceLab and consists of 263,424 utterances of Polish speech data from 200 speakers, totaling approximately 280 hours, and corresponding transcripts.

Data collection was performed in Poland. Speakers were asked to record themselves reading text on a website for at least 60 minutes from their home computer while using a headset. The read text was comprised of sentences covering most speech sounds in Polish.

This release includes speaker metadata. There were 103 male speakers and 97 female speakers, ranging from 15 – 60 years of age; most speakers were in the 15 – 30 years age range.

Polish Speech Database is distributed via web download.

2019 Subscription Members will automatically receive copies of this corpus. 2019 Standard Members may request a copy as part of their 16 free membership corpora. Non-members may license this data for $3000.

*

(3) 2016 NIST Speaker Recognition Evaluation Test Set was developed by LDC and NIST (National Institute of Standards and Technology) and contains approximately 340 hours of short segments of Tagalog, Cantonese, Cebuano, and Mandarin telephone speech used as development and test data in the NIST-sponsored 2016 Speaker Recognition Evaluation (SRE).

As in previous evaluations, SRE16 focused on telephone speech recorded over a variety of handset types for the training and test conditions. In addition to development and evaluation data, this corpus also contains trial lists, their associated keys, tables containing metadata information, and evaluation documentation.

The telephone speech data was drawn from the Call My Net 2015 Corpus collected by LDC. Native speakers of Tagalog, Cantonese, Cebuano, or Mandarin (220 unique speakers) made a total of ten telephone calls each to people within their existing social networks. Speakers were encouraged to use different telephone instruments in a variety of acoustic settings and were instructed to talk for 8 - 10 minutes per call on a topic of their choice. All conversations were collected outside North America.

2016 NIST Speaker Recognition Evaluation Test Set is distributed via web download.

2019 Subscription Members will automatically receive copies of this corpus. 2019 Standard Members may request a copy as part of their 16 free membership corpora. Non-members may license this data for $750.

*

Membership Office

Linguistic Data Consortium

University of Pennsylvania

T: +1-215-573-1275

E: ldc@ldc.upenn.edu

M: 3600 Market St. Suite 810

      Philadelphia, PA 19104

Top

Forensic database of voice recordings of 500+ Australian English speakers

We are pleased to announce that the forensic database of voice recordings of 500+ Australian English speakers is now published.

The database was collected by the Forensic Voice Comparison Laboratory, School of Electrical Engineering & Telecommunications, University of New South Wales as part of the Australian Research Council funded Linkage Project on making demonstrably valid and reliable forensic voice comparison a practical everyday reality in Australia. The project was conducted in partnership with: Australian Federal Police,  New South Wales Police,  Queensland Police, National Institute of Forensic Sciences, Australasian Speech Sciences and Technology Association, Guardia Civil, Universidad Autónoma de Madrid.

The database includes multiple non-contemporaneous recordings of most speakers. Each speaker is recorded in three different speaking styles representative of some common styles found in forensic casework. Recordings are recorded under high-quality conditions and extraneous noises and crosstalk have been manually removed. The high-quality audio can be processed to reflect recording conditions found in forensic casework.

The database can be accessed at: http://databases.forensic-voice-comparison.net/

Top

Audio and Electroglottographic speech recordings from several languages

We are happy to announce the public availability of speech recordings made as part of the UCLA project 'Production and Perception of Linguistic Voice Quality'.

http://www.phonetics.ucla.edu/voiceproject/voice.html

Audio and EGG recordings are available for Bo, Gujarati, Hmong, Mandarin, Black Miao, Southern Yi, Santiago Matatlan/ San Juan Guelavia Zapotec; audio recordings (no EGG) are available for English and Mandarin. Recordings of Jalapa Mazatec extracted from the UCLA Phonetic Archive are also posted. All recordings are accompanied by explanatory notes and wordlists, and most are accompanied by Praat textgrids that locate target segments of interest to our project.

Analysis software developed as part of the project – VoiceSauce for audio analysis and EggWorks for EGG analysis – and all project publications are also available from this site. All preliminary analyses of the recordings using these tools (i.e. acoustic and EGG parameter values extracted from the recordings) are posted on the site in large data spreadsheets.

All of these materials are made freely available under a Creative Commons Attribution-NonCommercial-ShareAlike-3.0 Unported License.

This project was funded by NSF grant BCS-0720304 to Pat Keating, Abeer Alwan and Jody Kreiman of UCLA, and Christina Esposito of Macalester College.

Pat Keating (UCLA)

Top

We are making 24 GB of a new dataset, called Kara One, freely available. This database combines 3 modalities (EEG, face tracking, and audio) during imagined and articulated speech using phonologically-relevant phonemic and single-word prompts. It is the result of a collaboration between the Toronto Rehabilitation Institute (in the University Health Network) and the Department of Computer Science at the University of Toronto.

In the associated paper (abstract below), we show how to accurately classify imagined phonological categories solely from EEG data. Specifically, we obtain up to 90% accuracy in classifying imagined consonants from imagined vowels and up to 95% accuracy in classifying stimulus from active imagination states using advanced deep-belief networks.

Data from 14 participants are available here: http://www.cs.toronto.edu/~complingweb/data/karaOne/karaOne.html.

If you have any questions, please contact Frank Rudzicz at frank@cs.toronto.edu.

Best regards,

Frank

PAPER Shunan Zhao and Frank Rudzicz (2015) Classifying phonological categories in imagined and articulated speech. In Proceedings of ICASSP 2015, Brisbane Australia

ABSTRACT This paper presents a new dataset combining 3 modalities (EEG, facial, and audio) during imagined and vocalized phonemic and single-word prompts. We pre-process the EEG data, compute features for all 3 modalities, and perform binary classi?cation of phonological categories using a combination of these modalities. For example, a deep-belief network obtains accuracies over 90% on identifying consonants, which is signi?cantly more accurate than two baseline supportvectormachines. Wealsoclassifybetweenthedifferent states (resting, stimuli, active thinking) of the recording, achievingaccuraciesof95%. Thesedatamaybeusedtolearn multimodal relationships, and to develop silent-speech and brain-computer interfaces.

Top

In the spirit of the season, I would like to announce the immediate availability of the TORGO database free, in perpetuity for academic use. This database combines acoustics and electromagnetic articulography from 8 individuals with speech disorders and 7 without, and totals over 18 GB. These data can be used for multimodal models (e.g., for acoustic-articulatory inversion), models of pathology, and augmented speech recognition, for example. More information (and the database itself) can be found here: http://www.cs.toronto.edu/~complingweb/data/TORGO/torgo.html.

Top

Datatang is a global leading data provider that specialized in data customized solution, focusing in variety speech, image, and text data collection, annotation, crowdsourcing services.

Top

Fearless Steps Corpus

John H.L. Hansen, Abhijeet Sangwan, Lakshmish Kaushik, Chengzhu Yu Center for Robust Speech Systems (CRSS), Eric Jonsson School of Engineering, The University of Texas at Dallas (UTD), Richardson, Texas, U.S.A.

NASA’s Apollo program is a great achievement of mankind in the 20th century. CRSS, UT-Dallas has undertaken an enormous Apollo data digitization initiative where we proposed to digitize Apollo mission speech data (~100,000 hours) and develop Spoken Language Technology based algorithms to analyze and understand various aspects of conversational speech. Towards achieving this goal, a new 30 track analog audio decoder is designed to decode 30 track Apollo analog tapes and is mounted on to the NASA Soundscriber analog audio decoder (in place of single channel decoder). Using the new decoder all 30 channels of data can be decoded simultaneously thereby reducing the digitization time significantly. 
We have digitized 19,000 hours of data from Apollo missions (including entire Apollo-11, most of Apollo-13, Apollo-1, and Gemini-8 missions). This audio archive is named as “Fearless Steps Corpus”. This is one of the most unique and singularly large naturalistic audio corpus of such magnitude. Automated transcripts are generated by building Apollo mission specific custom Deep Neural Networks (DNN) based Automatic Speech Recognition (ASR) system along with Apollo mission specific language models. Speaker Identification System (SID) to identify the speakers are designed. A complete diarization pipeline is established to study and develop various SLT tasks. 
We will release this corpus for public usage as a part of public outreach and promote SLT community to utilize this opportunity to build naturalistic spoken language technology systems. The data provides ample opportunity setup challenging tasks in various SLT areas. As a part of this outreach we will be setting “Fearless Challenge” in the upcoming INTERSPEECH 2018. We will define and propose 5 tasks as a part of this challenge. The guidelines and challenge data will be released in the Spring 2018 and will be available for download for free. The five challenges are, (1) Automatic Speech Recognition (2) Speaker Identification (3) Speech Activity Detection (4) Speaker Diarization (5) Keyword spotting and Joint Topic/Sentiment detection.
Looking forward for your participation (John.Hansen@utdallas.edu) 

Top

OPENGLOT –An open environment for the evaluation of glottal inverse filtering

OPENGLOT is a publically available database that was designed primarily for the evaluation of glottal inverse filtering algorithms. In addition, the database can be used in evaluating formant estimation methods. OPENGLOT consists of four repositories. Repository I contains synthetic glottal flow waveforms, and speech signals generated by using the Liljencrants–Fant (LF) waveform as an excitation, and an all-pole vocal tract model. Repository II contains glottal flow and speech pressure signals generated using physical modelling of human speech production. Repository III contains pairs of glottal excitation and speech pressure signal generated by exciting 3D printed plastic vocal tract replica with LF excitations via a loudspeaker. Finally, Repository IV contains multichannel recordings (speech pressure signal, EGG, high-speed video of the vocal folds) from natural production of speech.

OPENGLOT is available at:

http://research.spa.aalto.fi/projects/openglot/

Top

Nous sommes heureux de vous annoncer la publication d¹un ouvrage consacré
au treebank Rhapsodie, un corpus de français parlé de 33 000 mots
finement annoté en prosodie et en syntaxe.

Accès à la publication : https://benjamins.com/catalog/scl.89 (voir flyer
ci-joint)

Accès au treebank : https://www.projet-rhapsodie.fr/
Les données librement accessibles sont diffusées sous licence Creative
Commons.
Le site donne également accès aux guides d¹annotations.

Top

The My Science Tutor Children?s Conversational Speech Corpus (MyST Corpus) is the world?s largest English children?s speech corpus.  It is freely available to the research community for research use.  Companies can acquire the corpus for $10,000.  The MyST Corpus was collected over a 10-year period, with support from over $9 million in grants from the US National Science Foundation and Department of Education, awarded to Boulder Learning Inc. (Wayne Ward, Principal Investigator).

The MyST corpus contains speech collected from 1,374 third, fourth and fifth grade students.  The students engaged in spoken dialogs with a virtual science tutor in 8 areas of science.  A total of 11,398 student sessions of 15 to 20 minutes produced a total of 244,069 utterances.  42% of the utterances have been transcribed at the word level.  The corpus is partitioned into training and test sets to support comparison of research results across labs. All parents and students signed consent forms, approved by the University of Colorado?s Institutional Review Board,  that authorize distribution of the corpus for research and commercial use. 

The MyST children?s speech corpus contains approximately ten times as many spoken utterances as all other English children?s speech corpora combined (see https://en.wikipedia.org/wiki/List_of_children%27s_speech_corpora). 

Additional information about the corpus, and instructions for how to acquire the corpus (and samples of the speech data) can be found on the Boulder Learning Web site at http://boulderlearning.com/request-the-myst-corpus/.   

Top

• HARVARD speech corpus - native British English speaker, digital re-recording

Top