Please refer to the guide:

NVIDIA AudioEffects SDK

Prerequisites

AudioEffects SDK requires an NVIDIA Graphics driver >= 520.61.05, and is tested in Ubuntu 18.04/20.04, Debian 10 and RHEL/CentOS 7+.

Previous versions of drivers can be used by using the CUDA Forward-Compatible Update (refer to Section 3.2.1: Using Older Drivers(450/470) with CUDA Forward-Compatible Upgrade) in the Programming guide for more details).

All other libraries required for the SDK are already bundled under external/cuda/lib, do not need to be expliticly installed.

Installing

Installing NVIDIA Graphics Driver

Please refer to the installation guide at https://docs.nvidia.com/datacenter/tesla/tesla-installation-notes/index.html

Installing the SDK

Extract the SDK to the desired location on your computer, for example, by using the following command:

tar xvf Audio_Effects_SDK.tar.gz

Software dependencies

The SDK depends on CUDA 11.8, TensorRT 8.5.1.7 and CuDNN 8.6.0.

Please note that all dependent libraries are already included under external/cuda/lib, do not need to be explicitly installed.

The sample programs using the SDK reference these libraries via RPATH. In case the folder structure changes, LD_LIBRARY_PATH/similar can be used to explicitly load these libraries. For example

export LD_LIBRARY_PATH=/maxine_sdk_install/external/cuda/lib:$LD_LIBRARY_PATH

# This command will first check for libraries under /maxine_sdk_install/external/cuda/lib
./run_effect.sh -g a100

For further details, please refer to your distribution's documentation.

Running the sample programs

The SDK provides two sample apps, effects_demo and effects_delayed_streams_demo, both as pre-built binaries and with source.

To run the effects_demo sample on an A2 GPU, cd to samples/effects_demo and run the sample program using the run_effect.sh helper script, for example

./run_effect.sh -g a2 -e denoiser -s 48 -b 10

runs the 48k denoiser effect for batch size 10 with the model for A2 (sm_86 model).

To run the sample on all WAV files in a folder/specific WAV file, run the program using the following command:

# Run on all wav files under the ~/test_input folder (equivalent to ~/test_input/*.wav)
./run_effect.sh -g a2 -e denoiser -s 48 -i ~/test_input

# Run only on /test/sample.wav
./run_effect.sh -g a2 -e denoiser -s 48 -i /test/sample.wav

For further details about the sample programs, please refer to Section 3.3: Sample Applications in the Programming Guide

Supported effects

The following effects are supported:

# Denoiser
./run_effect.sh -g a16 -e denoiser -s 16 -b 10
./run_effect.sh -g a16 -e denoiser -s 48 -b 10

# Dereverb
./run_effect.sh -g a16 -e dereverb -s 16 -b 10
./run_effect.sh -g a16 -e dereverb -s 48 -b 10

# Dereverb+Denoiser
./run_effect.sh -g a16 -e dereverb_denoiser -s 16 -b 10
./run_effect.sh -g a16 -e dereverb_denoiser -s 48 -b 10

# AEC
./run_effect.sh -g a16 -e aec -s 16 -b 10
./run_effect.sh -g a16 -e aec -s 48 -b 10

# Superres 8k-16k
./run_effect.sh -g a16 -e superres -s 8 -o 16 -b 10
# Superres 16k-48k
./run_effect.sh -g a16 -e superres -s 16 -o 48 -b 10
# Superres 8k-48k
./run_effect.sh -g a16 -e superres -s 8 -o 48 -b 10

The following chained effects are supported:

# Denoiser 16k + Superres 16k-48k
./run_effect_chained.sh -e1 denoiser -s1 16 -e2 superres -s2 16 -g a16 -b 3
# Dereverb 16k + Superres 16k-48k
./run_effect_chained.sh -e1 dereverb -s1 16 -e2 superres -s2 16 -g a16 -b 3
# Dereverb+Denoiser 16k + Superres 16k-48k
./run_effect_chained.sh -e1 dereverb_denoiser -s1 16 -e2 superres -s2 16 -g a16 -b 3
# Superres 8k-16k + Denoiser 16k
./run_effect_chained.sh -e1 superres -s1 8 -e2 denoiser -s2 16 -g a16 -b 3
# Superres 8k-16k + Dereverb 16k
./run_effect_chained.sh -e1 superres -s1 8 -e2 dereverb -s2 16 -g a16 -b 3
# Superres 8k-16k + Dereverb+Denoiser 16k
./run_effect_chained.sh -e1 superres -s1 8 -e2 dereverb_denoiser -s2 16 -g a16 -b 3

Building applications with the SDK

For sample workflow for building applications with the SDK, please refer to Section 4.1: End to End flow in SDK application

v1.3.0

Latency for Superresolution effect has been reduced. The maximum batch size for these effects has also changed.