Freeing Speech Common Voice and Deep Speech Why

Free(ing) Speech Common Voice and Deep Speech

Why? Common Voice Deep Speech

Why?

How many languages have production quality, open speech recognition models?

Just one, English.

Why?

Despite the existence of various open STT engines ZERO LANGUAGES have the 10 k hours of open data needed for a production quality model.

Libri. Speech The largest open English corpus is Libri. Speech which is about 1 k hours

Formosa Grand Challenge Corpus The largest open Mandarin corpus is the Formosa Grand Challenge corpus which is about 400 hours

For Celtic languages. . . The availability of data sets of any size, free or at cost , drops off significantly

Common Voice

Collect

Validate

Distribute

And there’s more. . .

And there’s still more. . .

Deep Speech

Simple

A design principle we decided upon from the beginning is that the engine should work for ALL LANGUAGES the only requirement being that one have training data.

Perfection is achieved, not when there is nothing more to add, but when there is NOTHING LEFT TO TAKE AWAY -Antoine de Saint. Exupéry

More controversially, we decided that training a new language should not require LINGUISTIC KNOWLEDGE

Deep Speech Architecture Softmax Layer Feedforward Layer Recurrent Layer Feedforward Layers Input Features

Open

Deep Speech source code is released under Mozilla Public License 2. 0

Deep Speech modelsare released under Mozilla Public License 2. 0

Ubiquitous

Currently we support nine different Programming Languages

At the same time we also support nine different Platforms

Why? Common Voice Deep Speech

Why? Question: How many languages have production quality, open speech recognition models? Answer: Just one, English.

Common Voice ● Collect ● Validate ● Distribute

Deep Speech ● Simple ● Open ● Ubiquitous

Free(ing) Speech