NVIDIA Maxine Windows Audio Effects SDK
This is my tentative workflow for cleaning-up poor audio using the NVIDIA Maxine Windows Audio Effects SDK.
Download (requires NVIDIA developer account)
NVIDIA Maxine Documentation
The Audio Effects SDK only accepts audio that is in .wav format sampled at 8000Hz single-channel, 16000Hz single-channel, or 48000Hz single-channel. This can be resampled using ffmpeg.
- Convert a .mp3 file to 8kHz, single-channel PCM:
1ffmpeg -i input.mp3 -ar 8000 -ac 1 8000.wav
- Convert a .m4a file to 16kHz, single-channel PCM:
1ffmpeg -i input.mp3 -ar 16000 -ac 1 16000.wav
- Convert a .wav file to 48kHz, single-channel PCM:
1ffmpeg -i input.mp3 -ar 48000 -ac 1 48000.wav
A config file has to be fed to a batch script. The "effects_demo" includes sample config files for different GPU architectures. I hava an NVIDIA RTX 4080, so I would customize the "denoiser48k_cfg_ada.txt" config file and run it with the "run_denoiser_48k_ada.bat" batch file. Example config:
1# Effect.
2# Supported values are: denoiser/dereverb/dereverb_denoiser/aec/superres
3effect dereverb_denoiser
4# Point this to the model file.
5# This indicates 48k model for denoiser effect for ADA supported GPU architecture is picked from models folder
6# Similarly, this path can be modified as per user's choice of effect and sample rate (8k/16k/48k depending on effect)
7model models\ada\dereverb_denoiser_48k.trtpkg
8# Noisy input file
9# 48k Input file is picked from denoiser folder.
10# User can modify below line to pick their own file as input.
11input_wav input_files\48000.wav
12# Denoised audio data will be saved to this file.
13# Output can be dumped at user specifid location too. In this case, Output will be saved to current folder.
14output_wav 48000.wav
15# Set to 1 for real time mode i.e. audio data will be processed
16# at same speed like that of an audio input device like
17# microphone. Since the denoising is faster that real time, the
18# processing will be equal to audio file duration.
19real_time 0
20# Intensity Ratio
21intensity_ratio 0.5
22# Enable VAD
23enable_vad 1