Riva ASR Library - CUDA Feature Extractor

This is the Kaldi CUDA feature extractor. (Pulled out as a small, modular, standalone library.)

Build

Docker Images

• You can use one of these Kaldi containers from Nvidia NGC.
• It has been tested with nvcr.io/nvidia/kaldi:21.04-py3

To build libriva-asrlib-cuda-feature-extractor.a:

# From the repository root...
image="nvcr.io/nvidia/kaldi:21.04-py3"
docker image pull $image

docker run -e NVIDIA_VISIBLE_DEVICES=0 --runtime=nvidia \
    --user $(id -u):$(id -g) \
    -v /etc/passwd:/etc/passwd:ro \
    -v /etc/group:/etc/group:ro \
    --cap-add=SYS_ADMIN --workdir=/opt/kaldi/  --rm -it -v `pwd`:/opt/kaldi/ \
    $image  /bin/bash

mkdir build
cd build
cmake ..
make -j

Usage

Create the configuration object:

• include/cu-feat-extr/options/online-batched-feature-info.h
• Instantiate a OnlineBatchedFeatureInfo.
  • Use this overload OnlineBatchedFeatureInfo(RivaASRLibFeat_SpectralFeatureType feature_type_in, bool use_cmvn_in)
  • This is the only config object that should be directly instantiated. It will automatically construct all sub-config objects.
• Modify the config member variables as desired.

Create the compute pipeline object:

• include/cu-feat-extr/cudafeat/online-batched-feature-pipeline-cuda.h
• Instantiate a OnlineBatchedFeaturePipelineCuda
• Use pipeline.GetMaxChunkFrames() to help resize and prepare IO variables for the compute call below.
• Call pipeline.ComputeFeaturesBatched(...)
• Sync on cudaStreamPerThread as necessary.

Notes

• The library efficiently supports any combination of (online, offline) X (Batch Size = 1, Batch Size > 1).
• The following will be deprecated in a future release:
  • The C API for the Cuda Matrix class: c-cu-matrix.h. Use cu-matrix.h directly.
  • This overload of the configuration constructor:
    • OnlineBatchedFeatureInfo(const RivaASRLibFeat_OnlineBatchedFeatureInfo &opts_in)
• Syntax for configuration objects may change in a trivial way in a future release.
• Symbols in the following files will be put into a cpp namespace in a future release:
  • cu-container-helper-types.h
  • pipeline-helper-types.h
• There is limited support and testing for compatibility with Librosa.
  • For configurations that use frame_extraction_options.librosa_center = true, future releases of the library may produce slightly different output (frame count and frame content) or have a different interface to manage the padding. These both apply especially for online mode.
  • The configuration variable fbank_options.librosa_compat may be split up into multiple fine grain variables in a future release.
• Forked from Kaldi around commit 7f57eaa08093ba148e3d2abdf6c337212c130214