Library developers vosk Published New Models For Russian speech recognition: Server VOSK-model-EN-0.22 and mobile VOSK-Model-Small-RU-0.22. New speech data is used in models, as well as a new neuro-net architecture, which made it possible to increase the recognition accuracy by 10-20%. Code and data are distributed under the APACHE 2.0 license.
Important changes:
- New data collected in voice columns will significantly improve the recognition of speech commands sold from the distance.
- The new sound extraction scheme made it possible to significantly improve the recognition accuracy for broadband records. At the same time, the accuracy of telephony recognition also improved.
- Package for supplementing the dictionary allows you to customize the recognition of complex technical records.
For best accuracy, it is recommended to update the wax version to 0.3.32 . New wax features can also be interesting – integration with Unity, NativeScript, Jigasi. Models for recognition of Kazakh and Ukrainian languages. The server model for the work is needed a modern processor and 8GB of memory. Mobile model can be used in phones and RASPBERRYPI 3 +.
/Media reports.