The company Mozilla released the first issue of a speech recognition model of open source and a set of voice data, which included voice samples almost 20,000 people worldwide. Representatives of the company wrote in a blog project November 29, 2017.
The base of voice templates was collected thanks to volunteers: they dictated almost 400 thousand records with a total duration of 500 hours. All these records can be downloaded for free access. Such an available set will train the model for the machine learning system to the level of speech recognition. In this case, the number of errors will correspond to human recognition.
“ We believe that this technology will push to a wave of innovative products and services. It should be accessible to all, “- written in the company’s blog.
Now a text data set LibriSpeech is used, in which the number of recognition errors is only 6.5%. For the basis of the speech recognition system Mozilla uses the open source engine DeepSpeech.
More on: Popular solutions and developments in the field of voice technology
Having a model of speech recognition and examples of pronunciation, the developers will be able to experiment. Moreover, Mozilla offered ready-made modules for Python and NodeJS. With their help, you can embed speech recognition functions into your programs. There is also a toolkit for recognition from the command line.
While only English is supported, but in the plans of the creators in 2018 to begin collecting examples in other languages.