Fig. 1. The waveform, pitch and energy contour, word alignment, and sentence boundaries (denoted “SU”) for the utterance “um no I hadn’t heard of that”.

Fig. 2. An example of a decision tree for SU detection. Each line represents a node in the tree, “with the associated question regarding one particular prosodic feature, the class distribution, and the most likely class among the examples going through this node (S for SU boundary, and 0 for non-SU boundary). The indentation represents the depth of the nodes in the decision tree. The features used in this tree are described in Table 1.

Fig. 3. The bagging algorithm. T is 50 in our experiments. In each bag, the class distribution is the same as in the original data S.

Fig. 4. ROC curves for the decision trees trained from different sampling approaches and the original training set.

Fig. 5. ROC curves and their AUCs for the decision trees when bagging is used on the downsampled training set (bag-ds), the ensemble downsampled training sets (bag-ensemble), and the original training set (bag-original).

Examples of the prosodic features used for the SU detection problem

The “word” in the feature description is the word just before the boundary.

Description of the data set used in the pilot study

Note that the number of words shown in this table may not be exactly the same as that in the released data due to our processing of the data (e.g., text normalization, forced alignment).

Experimental results (CER in % and F-measure) for different sampling approaches in the SU detection task in the pilot study, using the prosody model alone and its combination with the LM

The CER of the LM alone on the test set is 5.02%.

Recall and precision results (%) for the sampling methods in the pilot study

Using the LM alone yields a recall of 74.6% and precision of 84.9%.

Experimental results (CER in % and F-measure) with bagging applied to a randomly downsampled data set, ensemble of downsampled training sets (DS), and the original training set

The results for the training conditions without bagging are also shown for comparison. The CER of the LM alone on the test set is 5.02%.

Information about CTS and BN corpora for the SU detection task

The speech recognition output is from SRI’s recognizer (Stolcke et al., 2000).

Experimental results (SU error rate in %) for both CTS and BN tasks on the REF and STT conditions, using the prosody model alone, and in combination with the LM

Three conditions are evaluated for the prosody model, random downsampling (prosody-ds), bagging on ensemble downsampling (prosody-ens-bag), and bagging on the original training set (prosody-bag-original).