Phonetics-based approaches allow for a high speed spotting and OOV detection, but the system's performance suffers from a lack of linguistic information that help distinguish targeted terms from phonetically close utterances [19], especially on short phonetic sequences [11].
Figure5A shows the feature vectors of close-talking utterance, which was recorded from the distance of 25 cm.
Thus, for these experiments, there were 150 pairs of distant-talking and close-talking utterances (1 session × 3 speakers × 10 utterances × 5 positions).
Although they were recorded from the distance of 25 cm, we regarded the close-talking utterances as our clean speech signals and tried to map the distant-talking utterances to close-talking utterances using NNs.
The non-stereo dataset contained close-talking utterances recorded from the distance of 25 cm and distant-talking utterances recorded from various positions in a room.
In this dataset, there were also close-talking utterances (25 cm), but they were not simultaneously recorded with the distant-talking utterances.
Both datasets consisted of close-talking utterances, which were recorded from the distance of 25 cm, and distant-talking utterances, which uttered from five different positions (P01 to P05) and captured by eight microphone arrays.
Table16 shows the upper bound and the baseline for set D. In the table, 'close' denotes close-talking utterances and represents the upper bound.
The GMMs for the SID system were trained using 10 close-talking utterances from the non-stereo dataset for each speaker.
For the experiments using real dataset, it is necessary to use CMN because we need to remove the noise and reverberation in the close-talking utterances used to train the speaker models.
Clinicians in our study used mainly closed questions (98/103 questions), made directive utterances ("Test your blood glucose on an empty stomach") and made statements about objective reality ("your glucose is too high").
