Jinay1991 Spleeter Versions Save

Release Note

Features

• Provides inferencing with TensorFlow C++ or TensorFlow Lite C++ APIs
• Provides precompiled binary of TF v2.3 or TFLite v2.3
• Provides TF Lite Model conversion script
• Provides Complete C++ Interface of https://github.com/deezer/spleeter

Known Issues

1. When running with TFLite Inference Engine, output of bass and accompaniment are swapped in resultant *.wav

CI Status

https://gitlab.com/jinay1991/spleeter/-/pipelines/210315716

Attachements

Results with TF and TFLite for given audio sample.

Pre-built binaries for TensorFlow and TensorFlow Lite were taken from https://github.com/jinay1991/perception/releases/tag/v0.5

Pre-trained Model checkpoints were taken from https://github.com/deezer/spleeter/releases/tag/v1.4.0 All models were converted using the scripts provided by https://github.com/tinoucas/spleeter-tflite-convert

2stems.tar.gz -> saved_model + tflite model for 2stems 4stems.tar.gz -> saved_model + tflite model for 4stems 5stems.tar.gz -> saved_model + tflite model for 5stems models.tar.gz -> contains all of the above models in one tarball.

Note: All the model binaries are from TF1.x. In this repository, TF2.3/TFLite2.3 C++ APIs were used for inference and works fine. Though new release will be made once TF2.x checkpoints are available from official repository.

Sample Audio Input: https://gitlab.com/jinay1991/spleeter/uploads/ded94b0b51d7021328491bac73d6c00a/audio_example.wav Results for TF/TFLite: 5stems_[tf|tflite]_output.tar.xz