Does Riva 1.8.0b0 automatically build both streaming and offline speech recognition models if the --offline flag is passed to riva-build?


It seems that this is now the case since I noticed the model names are now suffixed with *-streaming-offline if the --offline flag is passed.
I would like to confirm that since it will significantly simplify our model deployment process.

1 Like

I’m also curious about this. This release broke our offline inference due to new limits on offline inference.

To reproduce run Recognize() on any file > 900s (15 minutes). Looks like what has happened is they have now added true Offline mode, whereas I believe it was previously streaming (not sure I understand how this was called offline?) for batch/offline jobs with a large chunk size.

@rleary Is the way to get the previous “offline mode” to simply use the current streaming pertained citrinet model but with larger chunk sizes?

1 Like

I think that model name should probably not include streaming in the name. I will verify. You will need to generate RMIRs and deploy separate online and batch pipelines, as the models are optimized upon deployment for different scenarios.



This is what I get from building citrinet with --offline. Looks like streaming is still included in the model filenames generated on riva-deploy.