emOSA open speech archive
The project creates a web site which allows public or (in the case of ongoing research, restricted) access to all kinds of Hungarian speech data. The speech archive is designed to support the archival needs of research that relies on spoken materials (linguistics, anthropology, ethnography, etc.) with special emphasis on modern speech technologies, and in the long term, automatic speech recognition. Initially the archive will offer automated format detection, format conversion, voice activity detection, as well as experimental speaker and language identification software.
The archive focuses on both Hungarian and the minority languages spoken in Hungary (eg. different Romani dialects), including neighboring and related languages (majority or minority languages and dialects). Spontaneous speech recorded in natural conditions, noisy background or with inferior equipment is preferred, but acoustically clean materials are not excluded. Transcriptions, where available, will be included. The project is supported by an Infra2 grant from the Hungarian Academy of Sciences.
The following sub-projects are in progress:
- Automatic separation of speech and non-speech (voice activity detection, Judit Ács),
- Automatic language identification (Márton Makrai),
- Speaker identification (Katalin Pajkossy),
- Language Modeling (Dávid Nemeskey).