6.1.2 Additional Training

In additional training, acoustic models are created by multi-condition training, or a clean acoustic model is used as a base acoustic model. Aside from those are basically the same as multi-condition learning.

For example, In the case of a 16-mixture triphone model based acoustic model, first, prepare an audio data for additional training with the same procedure as multi-condition training. Then, similiar to the training after increasing the mixture count in multi-condition training (use HERest), perform the additional training to the audio data for additional training.

By additional training, even though the acoustic model originally used a different training data and underwent training, depending on the adaptive learning that used additional training data, the performance that is near to an additional training data that underwent multi-condition training from the beginning can be achieved.

Basically, to conform with the same method as using the multi-condition training, it is recommended to use a large amount of data for additional training. In case a large amount of data cannot be acquired, or if it is desired to shorten the calculation time of adaptive training, the MLLR/MAP adaptation in the next section is recommended.