Los Alamos Natl. Lab., CIC-3, MS B256, Los Alamos, NM 87545
The discriminant power of articulatory movements was evaluated for six subjects on a speaker-dependent continuous speech recognition task using a hidden-Markov-model-based speech recognition system. The articulatory measurements were gathered by means of electromagnetic articulography and describe the movement of small coils fixed to the speakers' tongue, jaw, and lower lip during the production of 108 German sentences. Four different articulatory representations were evaluated: coil displacements and their first three time derivatives (coil velocities, accelerations, and jerks). From these four representations, the coil accelerations performed by far the best in terms of recognition performance, both with and without acoustic features. The superior performance of acceleration features is surprising from the viewpoint of automatic speech recognition based on acoustics, since in the acoustic domain, acceleration features perform worse than static features on speaker-dependent tasks. From the viewpoint of articulatory phonetics, however, this result confirms the importance of the role articulator forces play in speech production.