Template adaptation in a hypersphere word classifier

Two methods of adapting speech recognition templates for developing multispeaker or speaker-independent vocabularies are tested. Both methods rely on feedback to assess the accuracy of recognition, and both use the test utterance to modify the correct template when recognition fails. In addition, one method uses the test utterance to modify intruding templates, moving them away from the test token's position in pattern space; this method is more effective for developing both multispeaker and speaker-independent vocabularies.<
>