Riva "unexpected shape for input 'audio_signal"

Please provide the following information when requesting support.

Hardware - GPU (T4)
Hardware - CPU
Operating System: Amazon Linux 2 (AWS ECS Optimized GPU Image)
Riva Version: 1.10.0b0

I’m trying to build a model using riva-build and riva-deploy commands following the instructions from the official documentation.

I am able to build models successfully, however, during deployment when I try to perform inference, I get the following error:

E0328 15:41:25.798205   645 grpc_riva_asr.cc:771] Triton client received error: in ensemble '<model-name>-streaming', unexpected shape for input 'audio_signal' for model 'riva-trt-<model-name>-streaming-am-streaming'. Expected [-1,80,460], got [1,80,461]
E0328 15:41:25.798382   236 grpc_riva_asr.cc:1116] ASRService.StreamingRecognize returning failure

This seems to occur for some configurations but not for others (Ex. chunk size, padding size).
Any idea why the input shape might be off by 1?

I was able to find certain configurations where Riva either returns no text back or the error above.
Everything else is default, or what was provided in the official docs:

Config Set 1:

  • chunk_size: 0.6
  • left_padding_size: 3.0
  • right_padding_size: 1.6, 2.4 (So either one of these basically)
  • vad_start_history: 300
  • vad_start_th: 0.2
  • vad_stop_history: 1600
  • vad_stop_th: 0.98

Config Set 2:

  • chunk_size: 0.7
  • left_padding_size: 3.0
  • right_padding_size: 2.4
  • vad_start_history: 300
  • vad_start_th: 0.2
  • vad_stop_history: 1600, 2000
  • vad_stop_th: 0.98

Config Set 3:

  • chunk_size: 0.8
  • left_padding_size: 3.0
  • right_padding_size: 0.7, 0.8
  • vad_start_history: 300
  • vad_start_th: 0.2
  • vad_stop_history: 1600
  • vad_stop_th: 0.98

Configs that do work however:

Config Set 1:

  • chunk_size: 0.5, 0.8
  • left_padding_size: 3.0
  • right_padding_size: 1.6, 2.4
  • vad_start_history: 300
  • vad_start_th: 0.2
  • vad_stop_history: 1600, 2000
  • vad_stop_th: 0.98

What is the reason for this?

Hi @pineapple9011 ,

Thanks for your interest in Riva,

I will check regarding this issue/error with the team and will provide an update soon,