Pattern Recognition Applications and Methods by Unknown

Author:Unknown
Language: eng
Format: epub
ISBN: 9783030054991
Publisher: Springer International Publishing

The 256 ms spectrogram windows typically can contain between 4 and 5 phones, with the average speaker uttering approximately 15 phones per second. The pooling layers in the CNN acoustic model provide flexibility in where the feature under question (phones in this case) can be within the 256 * 256 spectrogram image. This is useful for different orientations and scales of images in image classification and is also particularly useful for phone recognition where it is likely there will exist small errors in the training transcription.

During inferencing, the CNN acoustic model makes softmax predictions of all the phone classes for each of the test spectrograms, at three successive output stages of the network (Loss 1 to 3). We conducted some graphical analysis of the output confidences of the phones, colour coding the outputs for easier readability of the results, see Fig. 4. As can be seen from the loss-3 (accuracy), the network makes crisp classifications of usually only a single phone at a time. Given that this is unseen data, and that the comparison with the ground truth is good, we are confident that this network is an effective way to train an acoustic model.

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.