Skip to content

Roadmap

simonkingedinburgh edited this page Jun 14, 2017 · 9 revisions

Here's an attempt at roadmap for Merlin.

Tutorials

At some point, we might extend the stuff on speech.zone to be more general.

In the meantime, if you are coming to Interspeech 2017, then please come to our tutorial!

Using your own data, in English

Can be mostly handled by describing how to modify an existing recipe

Using another language

Will need to wait for Ossian to go on to GitHub, then add recipes to that, not Merlin.

Recipes

Blizzard Challenge benchmark systems

Add the 2016 benchmark system recipe. Need to change the dictionary to something Open Source, instead of Combilex.

2017 benchmark should use more sophisticated architecture - perhaps (simplified) LSTM and/or stacked bottleneck features.

Features

Unit and layer types

RELU units

Convolutional layers

Debugging, monitoring, and visualisation

Eye candy

Visualise t-SNE embeddings of hidden layer activations during training

Web interface to monitor training (extend the logger / plotter stuff)