dzdl/aicleaner

Audio denoising in real-time powered by artificial intelligence Python-friendly. Cross-platform. Check ROADMAP!

AiCleaner

VoiceCleaner
Audio denoising real-time powered by artificial intelligence (Beta)

GUI version: aicleaner gui

Name of the project may change on future.

Old-name: aicleaner

---

Table of Contents

• AiCleaner
  • Table of Contents
  • About The Project
    • How to run
      • On-premise Linux
      • Cloud hosted
      • Docker version
    • AI Models implemented
    • Roadmap
    • License

About The Project

This is an open-source project that wants to process audio input with denoising techniques, python-friendly and in real-time. All processing on-premise but can be handled by a server. Check the ROADMAP.

How to run

There are multiple ways to run: on-premise version (as personal user), on-cloud version (as enterprise digestor system) and docker version (as personal user in uncommon SO.). (suggest new one).

On-premise Linux

Apt

sudo apt-get update
sudo apt-get install tensorflow-model-server
# More info https://www.tensorflow.org/tfx/serving/setup

Pip

git clone https://github.com/DZDL/aicleaner
cd aicleaner
pip3 install -r requirements.txt
python3 main.py

On-premise Windows/MacOS/Linux (Cross-platform)

Needs Docker

# Tutorial comming

Cloud hosted

Not yet implemented.

Docker version

Not yet implemented.

AI Models implemented

AIModels supported by any ai-library on python (keras, tensorflow, pytorch, caffe). Create an issue if you want to add your model and we're going to adapt it.

Table 1. Models already implemented

✔️/x	Type	Model name	Repository	Demo	License
✔️	U-Net	Speech-Enhancement Max 8000 SR Port 8501	Original Adapted	Heroku	MIT
x	Unknown	DNP	Original Adapted	Broken	Unfound

Roadmap

Check the ROADMAP.

License

This open-source repository is under MIT license, check full here.

aicleaner's Issues

Dirty output audio

At the latest release working, a lot of noise can be hear.
May a filter can fix it, or increasing the time recording.
is this an error from the player, model, downsampling???

Cpython

Could a semi-compiled version improve the processing?

Tensorflow serving

One server is enough?
this method create parallel process for each input?

Error when playing

Need to check the algorithm and how it use numpy arrays, now 128 shape it's a problem

Best sound library in python for aicleaner

#	Name	SO	Record	Playback	Comment
1	pyalsaaudio	linux	✔️	✔️	It creates noise at the beginning of playing or record
2	soundcard	linux, macos, windows	✔️	✔️	working to handle input sequentially
3	sounddevice	linux,macos, windows	-	-	-
4	playsound	linux, macos, windows	-	-	-
5	simpleaudio	linux, macos, windows	-	-	-
6	winsound	windows	-	-	-
7	pydub	linux, macos, windows	✔️	✔️	problems with playing and recording due compatibility
8	pyaudio	linux, macos, windows	-	-	-
9	wavio	linux, macos, windows	-	-	too much delay when playing, at the beginning
10	soundfile	linux, macos, windows	-	-	-
10	pyalsaaudio	linux	✔️	✔️	some problems at the beginning, noise captured

More: https://realpython.com/playing-and-recording-sound-python/

GUI

GUI

Frontend required.

Don't write and read chunks

Improvement: pass the data only on RAM to decrease time processing and keep unstressed old hard disks

-Was needed in threading delegations.

Recording and player server is needed

Split on record and player servers the tasks in different threads to allow that
9b5964e

ADD CI/CD

in process

Error when playing with one_thread technique

Appears 1 inputs and 2 outputs created by ALSA.

However, after a period of time appears more and more outputs

Virtual mic

By now, it's only played on speakers, we need to create a virtual mic cross-platform to give users the ability to use that mic as mic filtered