Models

Sonosco provides some predefined deep speech recognition models, that can be used for training:

Deep Speech 2

We took the pytorch implementation of the Deep Speech 2 model from Sean Naren and ported it to the sonosco serialization guidelines.

Listen Attend Spell (LAS)

We took the pytorch implementation of the Listen Attend Spell model from AzizCode92 and ported it to the sonosco serialization guidelines. This model can be imported using: from sonosco.models import Seq2Seq

Sequence-to-Sequence Model with Time-Depth Separable Convolutions

We implemented a sequence-to-sequence model with Time-Depth separable convolutions in pytorch, following a paper from Facebook AI .

These models can be simply imported and used for training. (See Train your first model )