Prosodic event detection in children’s read speech

Deep Learning For Prominence Detection In Children's Read Speech

“…The children's oral reading dataset used in this work comprises recordings of grade-appropriate text read aloud by selected middle school students with reasonable word decoding ability in English (as second language) but widely varying levels of prosodic skill [11]. The individual utterances are story paragraphs comprising between 50-70 words, each word labeled separately for the presence/absence of prominence and phrase boundary by 7 naive listeners using the RPT methodology [29].…”

“…Performance of different architectures in single-task learning. A34 refers to the 34 hand-crafted features of [11]. Number of filters in all CNN and Sinc layers is 32 while pool size is 3.…”

“…We start with reporting the performance of the baseline system that uses a set of 34 word-level acoustic-prosodic features computed on the acoustic contours of pitch, intensity and spectral shape versus time as well as various segmental durations including pauses. The final set of 34 features (termed 'A34' here) is eventually obtained by two stages of feature selection in a random forest ensemble predictor [11]. We also test the same A34 set of features with a bidirectional GRU classifier to study whether the implicit 'learning' of word context can bring benefits.…”

“…These features are then used to train a conventional supervised classifier for the automatic detection, possibly in combination with lexico-syntactic information [9,10]. In our own recent work, we used feature selection on a large set of features, computed across the distinct suprasegmental attributes of speech, in a random forest ensemble predictor to derive a compact set of interpretable features for speaker-independent boundary and prominence detection on a children's oral reading dataset [11]. It was observed that apart from the expected pitch, duration and intensity based aggregates, the acoustic cues to prominence included a number of spectral shape functionals while the phrase boundary prediction was dominated by pause based features.…”

“…Our performance baseline is previous work on the same dataset and task in which a random forest ensemble predictor uses highly tuned hand-crafted features [11]. We also examine the possibly complementary information of the hand-crafted features.…”