Manual and automatic tone annotation: the case of an endangered language from north vietnam "mo piu"

The goal of our study is twofold. First, the results of two approaches in a tone annotation task are presented: a manual approach provided by the system MISTRAL+ and an automatic approach provided by the system PROSOTRAN. Second, an attempt is made here to determine the number and the patterns of the Mo Piu language tones with respect to the pitch slope and the tone level. It is found that both approaches of the tone annotation yield similar results on the speech material studied: tone patterns detected by the two systems are mainly falling (68% in the manual detection and 55% in the automatic detection) and flat tones (30% in manual detection and 43% in automatic detection) while rising tones are detected very seldom by the two systems (only 2% of the patterns are considered as rising ones). The agreement between the two systems is quite high: 70% of the detected tones have the same pitch movement (rising or flat) and only 4% of the tones detected by the two systems are completely different (neither the slope direction nor the tone levels are identical). The main tone patterns detected in our study are /41, 43, 54/ as falling slopes and /33/ as a plateau.