Second European Conference on Speech Communication and Technology

Genova, Italy
September 24-26, 1991


On the Structure of Subword Units for a Speaker Independent Continuous Speech Task

Isabel Galiano, Francisco Casacuberta, Emilio Sanchis

Departamento de Sistemas Informaticos y Computacion, Universidad Politecnica de Valencia, Valencia, Spain

The subword units in a continuous speech recognition system can be represented under different formalisms, but stochastic finite state networks are the most popular (i. e. hidden Markov models). However, the design of the structural component of these models requires a heuristic tailoring based on some a priori knowledge and/or by experimentation. Recently, two grammatical inference methods have been proposed to automatically obtain the structural component of the subword units from training speech data. These methods are based on the Error Correcting Grammatical Inference algorithm and the Morphic Generator Grammatical Inference methodology. In this paper we present speaker-independent experiments of acoustic-phonetic decoding in a continuous Spanish speech corpus. The models used are context-independent and context-dependent. They were obtained by using both grammatical inference algorithms. KEYWORDS: Acoustic-Phonetic Decoding, Grammatical Inference, Subword Modelling, Speech Recognition.

Full Paper

Bibliographic reference.  Galiano, Isabel / Casacuberta, Francisco / Sanchis, Emilio (1991): "On the structure of subword units for a speaker independent continuous speech task", In EUROSPEECH-1991, 675-678.