How to change ASR model in RIVA Sample App(with RASA NLU) demo

Hardware - GPU - V100
Hardware - CPU - Intel(R) Xeon(R) CPU E5-2690 v4 @ 2.60GHz
Operating System - Linux
Riva Version - 2.4.0

I am trying out the RIVA Sample App using RASA NLU - Virtual Assistant (with Rasa) — NVIDIA Riva and want to use the different ASR models in it.
How can we change the ASR model being used for speech to text in the Sample App?


Hi @varun.c57

Thanks for your interest in Riva,

I will check with the internal team and let you know


Hi @varun.c57

Thanks for waiting with us, I have some updates from the team

note that Step 1 in the Running the Demo section of the virtual-assistant-rasa sample app’s README file is

Start the Riva Speech Server, if not already done. Follow the steps in the Riva Quick Start Guide.

The script in the Riva Skills Quick Start resource folder is where you would change the ASR model. Keep the following in mind:

  1. Make sure that service_enabled_asr=true.
  2. By default, asr_acoustic_model=("conformer"). If you’re using the shell scripts to download a pre-trained ASR model with a different architecture, change this variable. The other possible values (enclosed in quotes and parentheses) are citrinet_1024, citrinet_256, jasper, and quartznet.
  3. Make sure to set riva_model_loc appropriately. You should end up with (or already have) the following subdirectories in $riva_model_loc: artifacts, models, and rmir.
  4. By default, use_existing_rmirs=false. This means that when you run, the script will download the pre-trained .rmir files, pipelines, and artifacts into the rmir, models, and artifacts subdirectories of $riva_model_loc. On the other hand, if you’ve customized any ASR models and passed the associated .riva files into riva-build to generate .rmir files, you can either run riva-deploy or to generate the pipelines (which get placed in the models subdirectory). In either of those cases, set use_existing_rmirs=true. Also make sure that your .rmir files are placed in a directory named rmir and that riva_model_loc is set to rmir’s parent directory.
  5. Once either or riva-deploy has finished running, run to start the Riva server.
  6. Enter the IP address and port (50051 by default) of the Riva server in asr_config["RIVA_SPEECH_API_URL"] in the virtual-assistant-rasa sample app’s file.

Hope this helps.