Jarvis Hindi model giving gibberish output

AI & Data Science Deep Learning (Training & Inference) Riva
#1

Hi, We are trying to deploy a Hindi Jarvis model for a live speech-to-text service [live streaming]. We have converted the Jarvis model from a Nemo Hindi model using the following steps:

image
image960×541 83.2 KB

AWS EC-2 : p2x.large.
GPU: Nvidia Tesla T4
Cuda version: 11.2
Nvidia driver version; 460.80

Note: In the second step, we have used 1.0.0-b.3 servicemaker because we are using Jarvis API of 1.0.0-b3 version.
Using the above steps we are able to convert the Nemo model to Jarvis models.
Conversion pipeline: Nemo → quartz.onnx → quartznet_asr.enemo → quartznet_asr.jmir → Jarvis models.
By giving the model location in the config file of jarvis api 1.0.0-b.3, we were able to start the model by running ./jarvis_start.sh. On testing with a sample hindi audio file, we are getting gibberish output.

Nemo model output and expected jarvis model output:

nemo
nemo643×592 175 KB

The current output of jarvis hindi model:

jarvis
jarvis785×596 18.5 KB

#2

Hi @namanveer2000 ,
Can you please help us with the model and reproducible script so that we can try debugging it.

Thanks!

#3

We have converted the Nemo model to ejrvs instead of enemo model using the nemo2jarvis pip package available in the jarvis_quickstart:1.3.0-beta resource. We went ahead with the remaining conversion process [ nemo->ejrvs->jmir->models ] and we used the latest servicemaker for the conversion [ jarvis-speech:1.3.0-beta-servicemaker ] and the gibberish output was resolved.

2 Likes
#4

@AakankshaS Please reply to this query, Or please raise Nvidia Internal bug.