The developers of the Vosk library have released new models for Russian speech recognition: the server model vosk-model-ru-0.22 and the mobile model Vosk-model-small-ru-0.22. These models use new speech data and a new neural network architecture, which has improved recognition accuracy by 10-20%. The code and data are distributed under the Apache 2.0 license.
Important changes:
- New data collected from voice speakers significantly enhances the recognition of voice commands spoken from a distance.
- The new sound extraction scheme has greatly improved the recognition accuracy for wideband recordings. At the same time, the accuracy for telephony recognition has also improved.
- The dictionary extension package allows you to configure recognition for complex technical recordings.
For the best accuracy, it is recommended to update to Vosk version 0.3.32. Also, new features of Vosk may be of interest — integrations with Unity, Nativescript, and Jigasi. Models for recognizing the Kazakh and Ukrainian languages are available. A server model requires a modern processor and 8GB of RAM. The mobile model can be used in phones and Raspberry Pi 3+.
Source: opennet.ru
