THINKTech Research Center

Non-profit Ltd.

Speech Technology

We have substantial experience in designing and implementing a speech recognition engine. Currently we focus on the LVCSR (Large Vocabulary Continuous Speech Recognition) of morphologically complex languages - such as Hungarian. Besides, we contributed to multilingual speech recognition systems and are experienced in using language independent modeling techniques.

We are maintaining and enhanceing a hihgly flexible time synchronous WFST /Weighted Finite State Transducer/ decoder. (This is the core of a speech recognition engine). The decoder is able to perform real-time recognition with a vocabulary size of 1 000 000 (one million!) on a normal PC. Moreover, the engine is able to operare with morpheme-like subword units which theoretically means an unlimited vocabulary size.

We are also maintaining and refineing a general purpose knowledge-source integration and optimization tool. It operates on weighted finite state transducer which are the extensions of finite state machines.

We have experiences in building various level of speech recognition models, either based on abstract, language specific knowledge or on statistics. In general, we prefer an suggest statistical methods in speech processing.

Other Competences

Although our main strength is speech technology, based on our academical an industrial background, we may assist in several type of other technological issues:

  • Telecommunications
  • Human-Computer Interfaces
  • Information Technology
  • Natural Language Processing
  • Know-how Teaching