9th Annual Conference of the International Speech Communication Association

Brisbane, Australia
September 22-26, 2008

A Non-Acoustic Approach to Crosslingual Speech Recognition Performance Prediction

Chen Liu, Lynette Melnar

Motorola, USA

Crosslingual acoustic modeling is an effective technique for building acoustic models in the absence of native training data. A small amount of native speech data is still needed for verifying the crosslingual models by running an actual recognition test. In some very stringent yet realistic situations, however, even the test data may not be available. We introduce an algorithm that objectively predicts the recognition performance of crosslingual acoustic models. This approach does not require conducting of actual speech recognition tests with target-language speech data; nor does it depend on any acoustic measurement techniques. The algorithm is based on a series of linguistic metrics characterizing the articulatory phonetic and phonological information of phonemes from both the target and source languages. It is useful both for validating crosslingual models for speech recognition applications, and for making database acquisition decisions that could prove very cost-beneficial.

Full Paper

Bibliographic reference.  Liu, Chen / Melnar, Lynette (2008): "A non-acoustic approach to crosslingual speech recognition performance prediction", In INTERSPEECH-2008, 2719-2722.