## Russian model for Vosk

Copyright 2017-2050 AC Technologies LLC

This is a Vosk model for Russian


## Changes in version 0.42

Extended vocabulary, retrained LM and RNNLM

Our audiobooks    %WER   4.5
Our calls         %WER  28.2
Silero audiobooks %WER  11.1
Silero calls      %WER  36.0
Silero youtube    %WER  19.5
Golos crowd       %WER   4.4
Sova devices      %WER  17.9


## Changes in version 0.22

Multistream architecture, golos dataset (https://github.com/sberdevices/golos)

Our audiobooks    %WER  5.74
Our calls         %WER 30.60
Silero audiobooks %WER 13.35
Silero calls      %WER 37.38
Silero youtube    %WER 20.73
Golos crowd       %WER  8.65
Sova devices      %WER 19.71


## Changes in version 0.10

No accuracy changes, reorganize layout to directly load the model from vosk-api

Our audiobooks    %WER  6.49
Our calls         %WER 36.24
Silero audiobooks %WER 16.04
Silero calls      %WER 40.69
Silero youtube    %WER 26.81
Golos crowd       %WER 28.87
Sova devices      %WER 29.22


## Changes in version 0.9

Added RNNLM rescoring


## Changes in version 0.8

More data and augmentations, should be better on phone calls


## Changes in version 0.7

Bigger tdnn-f model with specaugment, more epochs
Much better language model trained on huge amount of texts
