[TLT3.0][Jarvis] Fine tuning Quartznet produces garbled transcript

I have followed this notebook to fine tune the quartznet model from NGC on a specific accent of English. The transfer learning was completed and I was even able to export and deploy my fine-tuned Quartznet model to Jarvis.

I have changed my config.sh to load only asr_models from my exported location and the nlp_models still remains the same but when I login to jarvis_speech docker container and compare the actual models that were loaded before and after finetuning, NLP models and other VAD related models that support ASR transcriptions are also missing inside jarvis.

As a result of this, my streaming recognize requests are giving garbled transcripts even though for the achieved WER post fine-tuning my fine-tuned model is able to recognize some parts of the sentence correctly which it’s able to recorrect only at the end. PFB the results:

I have tried setting interim_results=False but still no luck! @SunilJB Please help me identify where am I going wrong ? Is it not possible to load rest NLP models and supporting ASR models from NGC and load only finetuned core ASR model into jarvis ?

PS: @shilpa.suresh I tried to do the same experiment and got stuck at the same point which made me switch from Jasper to Quartznet for fine tuning. Please do help me in case you were able to find a fix for that issue. Cheers!

Could you please share your asr_speech-to-text-deployment.ipynb and asr_speech-to-text-training.ipynb?
Please share config.sh too.

@Morganh PFB the files used in my experiment:


# Copyright (c) 2021, NVIDIA CORPORATION. All rights reserved.


# NVIDIA CORPORATION and its licensors retain all intellectual property

# and proprietary rights in and to this software, related documentation

# and any modifications thereto. Any use, reproduction, disclosure or

# distribution of this software and related documentation without an express

# license agreement from NVIDIA CORPORATION is strictly prohibited.

# Enable or Disable Jarvis Services




# Specify one or more GPUs to use

# specifying more than one GPU is currently an experimental feature, and may result in undefined behaviours.


# Specify the encryption key to use to deploy models


# Locations to use for storing models artifacts


# If an absolute path is specified, the data will be written to that location

# Otherwise, a docker volume will be used (default).


# jarvis_init.sh will create a `jmir` and `models` directory in the volume or

# path specified.


# JMIR ($jarvis_model_loc/jmir)

# Jarvis uses an intermediate representation (JMIR) for models

# that are ready to deploy but not yet fully optimized for deployment. Pretrained

# versions can be obtained from NGC (by specifying NGC models below) and will be

# downloaded to $jarvis_model_loc/jmir by `jarvis_init.sh`


# Custom models produced by NeMo or TLT and prepared using jarvis-build

# may also be copied manually to this location $(jarvis_model_loc/jmir).


# Models ($jarvis_model_loc/models)

# During the jarvis_init process, the JMIR files in $jarvis_model_loc/jmir

# are inspected and optimized for deployment. The optimized versions are

# stored in $jarvis_model_loc/models. The jarvis server exclusively uses these

# optimized versions.


# The default JMIRs are downloaded from NGC by default in the above $jarvis_jmir_loc directory

# If you'd like to skip the download from NGC and use the existing JMIRs in the $jarvis_jmir_loc

# then set the below $use_existing_jmirs flag to true. You can also deploy your set of custom

# JMIRs by keeping them in the jarvis_jmir_loc dir and use this quickstart script with the

# below flag to deploy them all together.


# Ports to expose for Jarvis services



# NGC orgs





# Pre-built models listed below will be downloaded from NGC. If models already exist in $jarvis-jmir

# then models can be commented out to skip download from NGC


### Punctuation model


### Jasper Streaming w/ CPU decoder, best latency configuration

# "${jarvis_ngc_org}/${jarvis_ngc_team}/jmir_jarvis_asr_jasper_english_streaming:${jarvis_ngc_model_version}"

### Jasper Streaming w/ CPU decoder, best throughput configuration

# "${jarvis_ngc_org}/${jarvis_ngc_team}/jmir_jarvis_asr_jasper_english_streaming_throughput:${jarvis_ngc_model_version}"

### Jasper Offline w/ CPU decoder

# "${jarvis_ngc_org}/${jarvis_ngc_team}/jmir_jarvis_asr_jasper_english_offline:${jarvis_ngc_model_version}"

### Quarztnet Streaming w/ CPU decoder, best latency configuration

# "${jarvis_ngc_org}/${jarvis_ngc_team}/jmir_jarvis_asr_quartznet_english_streaming:${jarvis_ngc_model_version}"

### Quarztnet Streaming w/ CPU decoder, best throughput configuration

# "${jarvis_ngc_org}/${jarvis_ngc_team}/jmir_jarvis_asr_quartznet_english_streaming_throughput:${jarvis_ngc_model_version}"

### Quarztnet Offline w/ CPU decoder

# "${jarvis_ngc_org}/${jarvis_ngc_team}/jmir_jarvis_asr_quartznet_english_offline:${jarvis_ngc_model_version}"

### Jasper Streaming w/ GPU decoder, best latency configuration

# "${jarvis_ngc_org}/${jarvis_ngc_team}/jmir_jarvis_asr_jasper_english_streaming_gpu_decoder:${jarvis_ngc_model_version}"

### Jasper Streaming w/ GPU decoder, best throughput configuration

# "${jarvis_ngc_org}/${jarvis_ngc_team}/jmir_jarvis_asr_jasper_english_streaming_throughput_gpu_decoder:${jarvis_ngc_model_version}"

### Jasper Offline w/ GPU decoder

# "${jarvis_ngc_org}/${jarvis_ngc_team}/jmir_jarvis_asr_jasper_english_offline_gpu_decoder:${jarvis_ngc_model_version}"













if [[ ! -z ${jarvis_ngc_team} ]]; then





# define docker images required to run Jarvis



# define docker images required to setup Jarvis


# daemon names




# Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved.
# TLT spec file for fine-tuning a previously trained ASR model (Jasper or QuartzNet) on the MCV Russian dataset.

  create_tensorboard_logger: true

  max_epochs: 50   # This is low for demo purposes

# Whether or not to change the decoder vocabulary.
# Note that this MUST be set if the labels change, e.g. to a different language's character set
# or if additional punctuation characters are added.
change_vocabulary: false

# Fine-tuning settings: training dataset
  manifest_filepath: ???
  sample_rate: 16000
  labels: [' ', 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', 'i', 'j', 'k', 'l', 'm',
           'n', 'o', 'p', 'q', 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', "'"]
  batch_size: 16
  trim_silence: true
  max_duration: 16.7
  shuffle: true
  is_tarred: false
  tarred_audio_filepaths: null

# Fine-tuning settings: validation dataset
  manifest_filepath: ???
  sample_rate: 16000
  labels: [' ', 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', 'i', 'j', 'k', 'l', 'm',
           'n', 'o', 'p', 'q', 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', "'"]
  batch_size: 16
  shuffle: false
  max_duration: 16.7

# Fine-tuning settings: optimizer
  name: novograd
  lr: 0.001

@Morganh speech-to-text-training.ipynb


Could you please check if speech-to-text-deployment.ipynb is correct? I did not see anything you run inside it.

@Morganh Sorry! I had copied the commands from the notebook and ran them from my terminal so there are no outputs in the notebook. I am sharing the log files here:

export.log (3.9 KB)


/opt/conda/lib/python3.6/site-packages/asr/scripts/export.pyc --config-path /specs/speech_to_text --config-name export.yaml exp_manager.explicit_log_dir=/results/quartznet/jarvis restore_from=/results/quartznet/finetune/checkpoints/finetuned-model.tlt encryption_key=********** export_format=JARVIS export_to=asr-model.ejrvs

There is no update from you for a period, assuming this is not an issue any more.
Hence we are closing this topic. If need further support, please open a new one.

I tried to understand your above comment but not yet. Could you give more detailed steps? Log is appreciated.
I would like to reproduce on my side.

For “garbled transcript”, in your speech_to_text_training.ipynb, I cannot see the output in the section “ASR Inference” or " ASR Inference using ONNX". Could you try to run it and attach the result?
Again, it is appreciated that you can provide more details for me to reproduce. Thanks a lot.

More, can you run the default speech_to_text_training.ipynb and speech-to-text-deployment.ipynb successfully?