Speech is a multimodal phenomenom at the perception and pro-
duction ends, and that includes the suprasegmental level of
speech. This paper focuses on the auditory-visual nature of lex-
ical tones, a suprasegmental unit of speech that characterises
tone languages. A multimodal corpus consisting of audio and
Optotrak recordings of 33 markers in the face and head was
recorded with 3 native speakers of Cantonese. The recorded tra-
jectories of the Optotrak markers were parameterized as poly-
nomial coefficients and used as input to Linear Discriminant
Analysis models for classification between the 6 Cantonese lex-
ical tones. Face and head motion were able to classify between
lexical tones with above-chance accuracy for each speaker in-
dividually and for all speakers combined. Other analyses were
carried out to determine which face regions and types of head
motion had a stronger influence of the lexical tone classification
accuracy, and the movement of the eyebrows and of the larynx
stood out.