We wanted to announce that Vosk speech recognition server now supports GPU decoding. With CUDA and NVIDIA cards it can accurately recognize up to 300 streams in parallel on a single machine in real-time. If you are using Vosk with Asterisk and look to process more data, try it out.
You need a modern GPU card (like GTX1080) and CUDA11.4 installed to try it.
You can run the server with a simple docker command:
docker run --runtime=nvidia -p 2700:2700 alphacep/kaldi-en-gpu
Then you can connect to the server as usual, for example, you can integrate Vosk Asterisk module or just forward audio through ARI. See the previous topic.