Contents

Source Codes

AMAAI has developed different AMT models

Spectrogram Reconstruction AMT

This model [11] uses spectrogram reconstruction to boost transcription accuracy.

source code

../../../_images/reconstruction_model.png

Fig. 2 Model architecture for spectrogram reconstruction AMT

ReconVAT

ReconVAT [12] is a semi-supervised AMT framework. This work is an extension of previous work by combining both virtual adversarial training (VAT) [13] and spectrogram reconstruction AMT [11].

source code
demo

../../../_images/ReconVAT_model.png

Fig. 3 Model architecture for ReconVAT

Understanding AMT using attention mechanism

We tried to understand AMT using additive attention mechanism.

source code
demo

../../../_images/attention_AMT.png

Fig. 4 Model architecture for ReconVAT

previous

Problem Formulation

next

VAE for Audio