Sentence examples for pair of utterances from inspiring English sources

We did preliminary experiments by using small NN training dataset, including '1s.1u' dataset, which contained only one pair of utterances.

We defined three kinds of dataset, i.e., '1s.1u' (1 pair of utterances, 1 speaker), '1s.5u' (5 pairs, 1 speaker), and '3s.15u' (15 pairs, 3 speakers).

By using the '6 NNs' configuration, the skip1 3-1-0 frame selection, and only one pair of utterances for the NN training data, we could reach 34.8% of ERR relative to the use of CMN.

From the evaluation using SID, we reached 26.0% and 34.8% of error rate reduction (ERR) relative to the baseline by using simulated and real data, respectively, by using only one pair of utterances for matched condition cases.

For example, the best ERR (regardless the frame selection) for RIR type of 'living room' by using only 1 pair of utterances and single NN reached 26.0%, while by using 15 pairs of utterances and multiple NN reached 62.6% (Table2).

From this pool of training data, '1u' (1 pair of utterances by 1 speaker), '5u' (5 pairs of utterances by 5 speakers), '10u' (10 pairs of utterances by 10 speakers), and '15u' (15 pairs of utterances by 15 speakers) NN training datasets were created.

Some pairs of utterances from set A were used as the NN training data.

Some pairs of utterances from the adaptation set of set D were used as the NN training data.

Meanwhile, by using 15 pairs of utterances and multiple NNs ('24 NNs') configuration, the ERR reached 62.6%.

The NN training datasets were created for each position and consisted of one or five pairs of utterances (utterance A0-A4) from each speaker.

For a very limited stereo data (e.g., one or five pairs of utterances), the use of modified multiple NNs configuration could improve the performance.