Hardware - GPU T4
Operating System - Ubuntu 18.04.6 LTS
Riva Version - 1.9.0
NeMo Version - 1.6.1
How to reproduce the issue ? (This is for errors. Please share the command and the detailed log here)
I have trained the STT En Conformer-CTC Large model using NeMo framework on my specific domain. It trained well, sample nemo transcribe output -
Audio in ./split_audio/inference_tests/chunk0.wav was recognized as: its something in the coaching dna all right so africknzaga held the top spot and the a p po all of last season weve had five different programs spend time as the number one team in the nation this year the road to the final four awfully wide open so were going to break down the the true contenders using the top three lines in
This output is correct. I then converted the nemo model to riva and ran the following command to build the rmir-
docker run --rm --gpus 0 -v /home/avik_pal/ASR/nemo_training/riva_models:/data nvcr.io/nvidia/riva/riva-speech:1.9.0-beta-servicemaker riva-build speech_recognition /data/conformer-ctc-trained.rmir /data/conformer-trained-decoder.riva \ --name=conformer-en-US-trained-asr-streaming \ --featurizer.use_utterance_norm_params=False \ --featurizer.precalc_norm_time_steps=0 \ --featurizer.precalc_norm_params=False \ --ms_per_timestep=40 \ --nn.use_onnx_runtime \ --vad.vad_start_history=200 \ --vad.residue_blanks_at_start=-2 \ --chunk_size=0.16 \ --left_padding_size=1.92 \ --right_padding_size=1.92 \ --decoder_type=greedy \ --language_code=en-US
also tried with parameter
--greedy_decoder.asr_model_delay=-1. There were no errors when building rmir and then deploying models. But when doing streaming inference with riva I get output like this -
it s so m e th ing in the co a ch ing d a ll r ight so a f ri ck and z a g a he l d the to p s p o t in the a p p o le a ll of la s t s e a s on we ve ha d f i ve di f f er ent p ro g ra m s s p en d t i m e a s the n u m b er on e t e a m in the n ation this y e ar the ro a d to the f in al f o ur a w f u ll y w i d e o p en so e we re go ing to b re a k do w n the the t r u e co n t en d er s u s ing the to p th re e l in es in
The character classes are correct. This seems to be some issue with output merging. Am I missing some parameters when doing riva build?
It would be really helpful if someone could guide me in resolving this.
Thanks and regards.