1The open prosody database OMProDat contains recordings of 40 five sentence passages, originally taken from the European SAM project, each read by 5 male and 5 female speakers of each language. The database contains both primary data, the recordings, and secondary data in the form of different annotation files. Currently the database contains recordings and annotations for five languages: Korean, English, French and Chinese plus a smaller subset for several languages. All the data are freely available on Ortolang (https://www.ortolang.fr/fr/accueil/).
2This corpus was created as a deliverable of the European Esprit project 2589 SAM (Speech Assessments and Methodology) and its follow-up project SAM-A.
Eurom1 contained, in particular, a series of 40 continuous and thematically connected fivesentence passages, intended to represent a clean version of the various types of speech which speech technology might be expected to deal with.
The passages were based on identical themes for the different languages, freely translated and adapted from the original English texts for the different languages. Two sample passages from the Eurom1-EN database are:
[T02] I have a problem with my water softener.
The water-level is too high and the overflow keeps dripping.
Could you arrange to send an engineer on Tuesday morning please?
It’s the only day I can manage this week.
I’d be grateful if you could confirm the arrangement in writing.
[T33] Hello, is that the telephone-order service?
There seems to have been some mistake.
I ordered a teddy bear from the catalogue and was billed for an electric lawnmower.
And I don’t even have a garden.
Would you put me through to the complaints department, please?
3The passages were originally recorded in the 1980’s for eleven European languages:
Danish, Dutch, English, French, German, Greek, Italian, Norwegian, Portuguese, Spanish and Swedish.
4A compatible speech database for East European languages was later recorded during the Copernicus project 1304, Babel, with similar recordings for Bulgarian, Estonian, Hungarian, Polish, and Romanian (Roach et al 1996).
5The continuous passages from the Eurom1 corpus in five languages: English, French, Italian, German, and Spanish, were re-used during the Esprit project Multext.
The recordings were provided with manually created annotation files for word labels and with automatically stylised f0 patterns using the Momel algorithm Hirst & Espesser (1993), Hirst (2007).
The database was published as the MULTEXT Prosodic Database Campione & Véronis (1998).
6A compatible version of the database for East European languages was produced as Multext-East Erjavec 2004.
7The original Eurom1 recordings were protected by copyright assigned to the different laboratories that produced the recordings. For details see http://www.phon.ucl.ac.uk/shop/eurom1.php. The database contained on 30 CDs is available for sale from the same address for £100.
8The Babel corpora are available from ELRA (http://catalog.elra.info.product_info.php) at €600 per language for researchers. ELRA members get a 50\% discount and ELRA membership costs €750 for non-profit-making organisations.
9The Multext Database is available from the same address for €100 for academic researchers. The Multext-East recordings are freely available from http://nl.ijs.si/ME/.
10In order to provide a more solid basis for the analysis of prosodic metrics, we decided to build an open multilingual prosodic database OMProDat, to be archived and distributed by Ortolang under an open database license.
11The aim of this database is to collect, archive and distribute recordings and annotations of directly comparable data from a representative sample of different languages representing different prosodic typological characteristics.
12As mentioned above, the passages of the different versions of the original Eurom1 corpus were typically read by only two or three speakers each.
This makes the corpus of limited use for the study of speaker variability.
13It was consequently decided to make new recordings of the corpus, with all 40 passages read by 10 speakers each.
14The first language recorded under these conditions was Korean Kim et al (2008). The original English version of the Eurom1 text was translated and adapted for Korean.
The texts in Korean alphabet were Romanized and also transcribed in SAMPA and IPA. 10 Seoul speakers (5 male and 5 female) took part in the recording session, all were Korean native speakers in their twenties, either undergraduate or graduate students of Seoul National University. Each speaker read all 40 passages.
15For prosodic annotation, the Momel algorithm was used (Hirst 2007) and the pitch targets obtained were manually corrected. The prosodic events were annotated in two ways: first, with the automatic annotation algorithm, INTSINT (Hirst 2007) and second, with manual labelling of prosodic units using just two tone labels (H and L).
16New recordings were made for English and French read by native speakers, as well as for English read by native speakers of French and for French read by native speakers of English (Herment et al 2012). The speakers were all 20-30 years old. All speakers were from monolingual families. The English speakers were recorded in Oxford and spoke Southern British English; the French speakers were recorded in Aix-en-Provence and spoke either a Southern or a standard variety of French, or something between the two.
17The originality of this corpus is that it provides recordings for both natives and non-native speakers, so as to allow comparative studies on L1 and L2 productions.
18Three groups of learners were recorded for each language, one group of native speakers on the one hand and two groups of non-native speakers, corresponding to the levels of the Common European Framework of Reference for Languages*, (CEFR**): classified respectively as independent users (level B1/B2) and proficient users (level C1/C2).
19The recordings are accompanied by TextGrid annotation files obtained semi-automatically from the sound and the orthographic transcription using the SPPAS alignment software (Bigi & Hirst 2012) using manual correction when necessary
20Prosodic annotation was also obtained using the Momel and INTSINT automatic annotation algorithms (Hirst 2007).
21More recently, recordings for Standard Mandarin Chinese (Ding & Hirst 2012) were made. The speakers were 10 Chinese native speakers: 5 female and 5 male.
Their ages ranged from 21 to 31 years old, and they were all postgraduate students and speakers of standard Chinese.
Each speaker read all 40 passages. The annotation of the recordings using SPPAS and Momel/INTSINT is currently in progress.
22A more limited set of data from a larger number of languages were also collected and distributed.
23The two paragraphs from the English Eurom1 corpus given as examples above were translated and recorded. This shared corpus is hosted by Ortolang (https://hdl.handle.net/11403/openprodat/v2) under the name OpenProDat as a part of the more general OMProDat database described here.
24These texts were transcribed in: Dutch, French, German, Italian, Arabic, Spanish, Finnish, Hungarian, Japanese and Thai. Each participant read both paragraphs, first in their mother tongue and then in each language that they felt able to read.
25Some files were manually transcribed and annotated with SPPAS (Bigi & Hirst 2012). These annotations are also freely available in the Ortolang Platform..
Flux RSS
