Databases of Spoken Speech

Speech databases are necessary for creation of speech recognition systems as they serve as resources for statistics of acoustic characteristics (speech databases) or for models of particular word context (text corpora). Reasonable part of our activities are devoted to the collection of large speech and text databases. We have participated on creation of several large database created within Eurpoean project or on the basis of bilateral commercial cooperations.

Wideband speech databases

As wideband speech, we understand mainly the data collected out of telephone networks when used sampling franquency is usually 16 kHz or higher. We have participated on European FP5 project within wich we have created large database of 600 speakers (Czech SPEECON) designed for training of recognizers developed for voice controlled consumer devices under various conditions. (office, public places, home environment, car). Database contains signals collected by 4 microphone of different quality and placement (close-, middle-, and far-talk microphones). This database is avialable via ELRA (http://www.elra.info) under number S0298.

In collaboration with Radboud University and Max Planck Intitute of Psycholinguistics in Nijmegen (Netherlands) the Nijmegen Corpus of Casual Speech has been created. It is the database of spontaneous and strongly informal speech collected in groups of 3 speaker within their informal communication. This data were collected for the purposes of further studies of spontaneous speech recognition as well as for studies at the linguistic level.

From the requirements concerning research activities in the field of speech enhancement and robust recognition in the car, the need of database of car speech and car noises arised as necessary requirement. Within cooperation with TEMIC SDS and later with Harman/Becker (Ulm, Germany) we have created large database of 1000 speakers from running car. Created database contains typical utterances related to voice control of various function in car environment, and additionally also phonetically rich material.

Telephone speech databases

Telephone speech databases are required due to frequent applications of speech recognition in telephone information systems. We have created two important databases: "Czech SpeechDat" - database of telephone speech from 1052 speakers which contains various types of utternaces including phonetically rich sentences and words and "CISLOVKY (NUMERALS)" - database of isolated and connected digit, 1227 speakers. These databases aer avialable via ELRA catalogue (http://www.elra.info) under numbers S0077 (Numerals) a S0094 (Czech SpeechDat).

Special speech database with evoked Lombard effect

The research in the field of Lombard effect (its analysis and normalization) has brought the requirement to create a database with sufficiently large level of Lombard effect. On the basis of this requirement we have created unique database with evoked Lombard effect.

Text corpora and lexica

Large text corpora contain typically necessary input information for training of statistical language models, but they can serve also as the resource of phonetically rich material used for further recording of speech databases. Large corpus of publicly available text on Internet was collected for this purpose. The second group of text databases is represented by lexica, giving complete overview of language vocabulary but also serving the information about word pronunciation or about the morphology of given word form. We have participated on creation of large Czech lexicon of LC-Star family within the comercial project LC-Star2.