Spleeter_Unofficial_TF20_MobileApp
Mobile app that consumes the TFLite version of Spleeter_TF2.0_Unofficial version and performs the audio source separation on the edge.
Update - Mar 29, 2020
As mentioned below, this mobile app helps in the generation of audio files but with considerable noises. I am working on adding a denoiser to reduce the noises. Please click on the following soundcloud links to hear the quality of audio files:
- Input Audio
https://soundcloud.com/vasanth-velayudham/aclassiceducation
- Vocal extract from mobile app
https://soundcloud.com/vasanth-velayudham/aclassiceducation-vocal-extract
- Denoised Vocal extract from mobile app [This is in progress and the code will be committed soon]
https://soundcloud.com/vasanth-velayudham/aclassiceducation-vocal-denoised
Update - Mar 23, 2020
Committed the mobile app that uses the TFLite model of SpleeterTF2.0_Unofficial project and generates the vocal version of the given input.
Model file to extract vocal part is present under the directory 'app/src/main/assets' and this project would require a separate model for each of the stems. Right now, it is supporting only vocal extraction.
As you can hear from the 'sample_audio_files' directory, output generated contains the 'vocal' portion of the input audio with considerable noise. As per analysis, this considerable noise is the result of model conversion into TFLite format and the processing of audio data in Android.
Right now, I am working towards removing the noise by feeding the output of the above model to 'denoiser' model built with Deep Learning. If Denoiser model succeeds, this project would acheive a pretty decent success.
Keep following this project to get further updates on the denoiser integration.
This repo contains the code for Android app that consumes the model generated by the below repository:
https://github.com/VVasanth/SpleeterTF2.0_Unofficial
This android app processes the audio file, generates the features, feeds the features to TFLite model, processes the output of model and generates the audio file from the processed output.
You can use the sample audio files present under the 'sample_audio_files' directory for processing. Please ensure to place the input audio file under the directory 'mnt/sdcard/audio-separator-input' of your emulator, so that code can process the audio file. Processed output would be generated under the directory 'mnt/sdcard/audio-separator-output'.
Right now, entire processing happens and the e2e flow is working. But the results contain lot of noises. I will be working towards optimizing the noises and improving the accuracy in the coming weeks.