ATREUS: a comparative study of continuous speech recognition systems at ATR

The authors describe ATREUS, an aggregation of a large variety of continuous speech recognition systems, forming the spoken input front-end of an interpreting telephony system. ATREUS includes the following phone models: discrete HMMs (hidden Markov models) with fuzzy vector quantization (VQ) and multiple codebooks; continuous mixture density HMMs; hidden Markov networks derived from the SSS (successive state splitting) algorithm; time-delay-neural networks; and fuzzy partition models. Its speaker modes involve speaker-dependent, speaker-independent, and speaker-adaptive techniques such as codebook mapping for VQ-HMMs, vector field smoothing for all types of HMMs, and neural network speaker mapping. A comparative study is given from the viewpoints of structure, constituent techniques, hardware implementation, and performance. ATREUS was evaluated for Japanese phrase speech recognition. A combination called ATREUS/SSS-LR had the best performance among the ATREUS systems.<>