Robust Spoken Document Retrieval Based on Multilingual Subphonetic Segment Recognition

This paper describes the development and application of a subphonetic segment recognition system for spoken document retrieval. Following from the development of an open-vocabulary spoken document retrieval system, where the retrieval process is accomplished in the symbolic domain by measuring the distance between the parts of subphonetic segment results from pattern recognition in the acoustic domain, the system proposed here performs matching based on subphonetic segment as more basic unit than the semantic unit. As such, the system is not constrained by vocabulary or grammar, and can be readily extended to multilingual tasks. This paper presents the proposed spoken document retrieval system including the proposed subphonetic segment recognition scheme, and evaluates the performance and feasibility of the system through experimental application to multilingual retrieval tasks.