Any Good Practice for Data ingestion for ASR?

I was Trying to find a way to get RIVA ASR working with AWS for transcription. The challenge I am currently facing is how to get real-time transcription working. I always get issue on either using gRPC or transforming the data.
I would like to know if anyone found an optimal way to do real-time transcription in AWS?

Below there is some thoughts:
AWS Kineses Video Data Stream → (some data transformation) → Riva ASR (AWS Batch EKS)

  • If I use Firehose + Lambda it becomes slow and batch processing.
  • If I create a new Batch EKS just for the data transformation it adds complexity among other challenges.

I might Edit the post as I go exploring and discovering new solutions.

Hi @Pappachuck_renan

Thanks for your interest in Riva

I will check regarding your request with the team and get back

Thanks

Hi @Pappachuck_renan

I have some inputs from the team

Please follow the steps from How to Deploy Riva at Scale on AWS with EKS — NVIDIA Riva for deploying Riva API on AWS

Questions from the team

  1. For EKS Cluster setting, have you looked into managedNodeGroups configs?
  2. Can you share what the setting you are using for Firehose and Lamda ?
  3. Can you give specific details on errors which you get with AWS during transciption ?

Please provide the above details, will get back to the internal team and provide updates

Thanks

Your example they used a Ingress through traefik and I assume they do transform the data on the cluster itself.
I have seen examples on TensorRT.
I was mostly curious about alternatives to what u present.
Ingress through a edge router is far from optimal.

So far I am just playing with ideas. I thought I could use Kinesis so I can have a more robust integration. Firehose is used for some ELT pipelines, pretty popular. lambda would just convert the encoding and send the job request.
I did not set anything yet. It doesnt feel right yet. I need to read more and see what can be done.

The ASR I cant get it working at all outside NeMo. I am following other post for that.