{"id":53722,"date":"2019-12-08T00:00:00","date_gmt":"2019-12-07T21:00:00","guid":{"rendered":"https:\/\/prohoster.info\/blog\/blog_prohoster\/kompaniya-mozilla-predstavila-dvizhok-raspoznavaniya-rechi-deepspeech-0-6"},"modified":"2020-02-18T14:01:39","modified_gmt":"2020-02-18T11:01:39","slug":"kompaniya-mozilla-predstavila-dvizhok-raspoznavaniya-rechi-deepspeech-0-6","status":"publish","type":"post","link":"https:\/\/prohoster.info\/en\/blog\/news\/kompaniya-mozilla-predstavila-dvizhok-raspoznavaniya-rechi-deepspeech-0-6","title":{"rendered":"In the first part, I briefly talked about the build that allows setting up a computer on which unRAID can run.","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"

Introduced<\/a><\/noindex> The release of the speech recognition engine developed by Mozilla DeepSpeech 0.6<\/a><\/noindex>, which implements the namesake speech recognition architecture, proposed by<\/a><\/noindex> researchers from Baidu. The implementation is written in Python using the TensorFlow machine learning platform and is distributed<\/a><\/noindex> is licensed under MPL 2.0. It supports operation on Linux, Android, macOS, and Windows. The performance is sufficient for using the engine on LePotato boards, Raspberry Pi 3, and Raspberry Pi 4.<\/p>\n

The package also includes for installation in WSL<\/a><\/noindex> pre-trained models, examples<\/a><\/noindex> sound files, and command line tools for recognition. Ready-to-use modules for Python, NodeJS, C++, and .NET are provided for integrating speech recognition functionality into applications (third-party developers have separately prepared modules for Rust<\/a><\/noindex> and Go<\/a><\/noindex>). The available model is only for the English language, but for other languages, you can train the system yourself using the<\/a><\/noindex> the instructions<\/a><\/noindex> voice data , collected by the Common Voice project.<\/a><\/noindex>DeepSpeech is significantly simpler than traditional systems while providing higher recognition quality in the presence of background noise. The development does not use traditional acoustic models and the concept of phonemes; instead, it employs a well-optimized machine learning system based on neural networks, allowing for the absence of separate components for modeling various distortions such as noise, echo, and speech features. <\/p>\n

The downside of this approach is that achieving quality recognition and training the neural network requires a large volume of diverse data, recorded in real conditions by different voices and amidst natural noise. <\/p>\n

The collection of such data is handled by the Mozilla-created project
\nCommon Voice , providing a validated dataset with 780 hours in<\/a><\/noindex>the English language , 325 in German, 173 in French, and 27 hours in Russian.<\/a><\/noindex>, 325 in German, 173 in French, and 27 hours in Russian. <\/p>\n

The ultimate goal of the Common Voice project is to accumulate 10,000 hours of recordings of various pronunciations of standard phrases of human speech, which will allow for an acceptable level of errors in recognition. In its current form, project participants have already recorded a total of 4,300 hours, of which 3,500 hours have been validated. For training the final English language model for DeepSpeech, 3,816 hours of speech were used, in addition to Common Voice, covering data from the LibriSpeech, Fisher, and Switchboard projects, as well as including about 1,700 hours of transcribed radio show recordings.<\/p>\n

When using the pre-trained English language model available for download, the error rate in DeepSpeech is 7.5% when evaluated with a test dataset. LibriSpeech<\/a><\/noindex>. For comparison, the error rate when recognized by humans is estimated<\/a><\/noindex> at 5.83%.<\/p>\n

DeepSpeech consists of two subsystems \u2014 the acoustic model and the decoder. The acoustic model uses deep machine learning methods to compute the probability of certain characters in the input sound. The decoder applies a beam search algorithm to transform the character probability data into a text representation.<\/p>\n

Key innovations<\/a><\/noindex> DeepSpeech 0.6 (branch 0.6 is not compatible with previous releases and requires code and model updates):<\/p>\n