Canary ouputs English for Arabic Speech

[BenoitWang]

Describe the bug

Hello, I am trying to infer Canary 1b for Arabic ASR with Riva quick start 2.18.0. According to the description it has already supported Arabic, but it outputs English instead of Arabic tokens.

Here's a partial extract of my config.sh

service_enabled_asr=true
service_enabled_nlp=false
service_enabled_tts=false
service_enabled_nmt=false

asr_language_code=("ar-AR")
asr_acoustic_model=("canary")
asr_acoustic_model_variant=("1b")
use_existing_rmirs=false

When I used the same config to infer Parakeet 1.1b_unified_ml_cs_concat and Parakeet 1.1b_unified_ml_cs_universal, they do output Arabic tokens, so I guess the issue is within the Canary model.

Any idea please?