Radio-ASR

This package is used to convert demodulated audio signals into text. It leverages the NeMo ASR model.

NOTE: The installation order matters here. Follow the instructions step-by-step

Installation For JetPack 4.6 (AirStack 0.5+)

Some of the pip builds for NLP pkgs need a C++ compiler. Building some of NeMo's ASR dependencies also requires a Rust compiler

Prerequisites

• Install git-lfs in order to access the large binary files in this repository

  sudo apt-get install git-lfs
  git lfs install
  

• Install Binary Dependencies of NVIDIA-built PyTorch package

  sudo apt-get install libopenblas-base libopenmpi-dev
  

Create Initial Environment

Note: Always add compilers when creating the initial environment: they have activation scripts to set the environment, so otherwise would have to deactivate and reactivate the environment.

conda create -n gnuradio-asr -c file:///opt/deepwave/conda-channels/airstack-conda python=3.6 compilers rust numpy scipy matplotlib mamba
conda activate gnuradio-asr
mamba update mamba
mamba install -c file:///opt/deepwave/conda-channels/airstack-conda gnuradio soapysdr-module-airt
mamba install scikit-learn onnx ipython pandas notebook numba click=7 cython h5py \
              sympy editdistance nltk grpcio markdown werkzeug tensorboard=2.4

Install PyTorch

See documentation here for installing PyTorch on JetPack.

• The steps for PyTorch 1.9.0 (downloading the package from NVidia)

  wget https://nvidia.box.com/shared/static/h1z9sw4bb1ybi0rm3tu8qdj8hs05ljbm.whl
  mv h1z9sw4bb1ybi0rm3tu8qdj8hs05ljbm.whl torch-1.9.0-cp36-cp36m-linux_aarch64.whl
  pip install torch-1.9.0-cp36-cp36m-linux_aarch64.whl
  

  Note that the torch wheel is saved to this repository just in case it stops being published. You can alternately just run:

  pip install pkgs/torch-1.9.0-cp36-cp36m-linux_aarch64.whl
  

Install NeMo

• Clone the repository

  git clone https://github.com/NVIDIA/NeMo
  cd NeMo
  git checkout v1.3.0
  

• Basic requirements for NeMo

  cd NeMo/requirements
  pip install -r requirements.txt
  pip install -r requirements_asr.txt
  

• The model depends on a c++ package that's in conda, but not built for linux-aarch64. We could clone the feedstock and build it ourselves, but it also doesn't build cleanly. It's for Japanese language support which we don't need right now so we remove it.

  cd NeMo
  patch -p1 < ../nemo-1.3.patch
  

• The NLP code installs another package that has a C++ dependency that we don't have, so replace it with the pure python version of the same package here...

  pip uninstall opencc
  pip install opencc-python-reimplemented
  

• Add NeMo to our environment

  pip install ./NeMo
  

• One last step: build/install the external libraries necessary to run the beam search decoders & language models.

  cd NeMo
  scripts/asr_language_modeling/ngram_lm/install_beamsearch_decoders.sh
  

Install Radio-ASR

Run this command from the folder containing this readme (i.e. the top of this git repo)

pip install -e .

Now you SHOULD be done and can run the transcription examples. Whew!