AI-powered English learning apps are used by hundreds of millions of people across the globe on a daily basis. This presents a great opportunity for the study of L2 speech. On one hand, the amount of data accessible for research is very large and rapidly growing; on the other hand, new theories and understanding of L2 speech can be continually tested and revised through real-life and real-time applications. This paper presents a study of pitch characteristics of L2 English speech using a large-scale dataset from a language learning app. Our dataset contains 180,000 spoken utterances which amount to 240 hours of speech. The results show that compared to L1, L2 English has narrower pitch range and slower rate of pitch change, but more small "ripples" on the pitch contour. The percentage of F0 rise time is higher in L2, and the maximum F0 in an utterance is realized later (with respect to the onset of the word on which the maximum F0 resides). These results suggest that the influence of L1 on L2 prosody is more complex than previously demonstrated, and they shed light on L2 prosody assessment and learning.

No abstract

No abstract

No abstract