JusperLee

Kai Li 1,2 , Yi Luo 2 1 Tsinghua University, Beijing, China 2 Tencent AI Lab, Shenzhen, China ArXiv | Demo Apollo: Band-sequence Modeling for High-Quality Music Restoration in Compressed Audio Apollo is a novel music restoration method designed to address distortions and artefacts caused by audio codecs, especially at low bitrates. Operating in the frequency domain, Apollo uses a frequency band-split module, band-sequence modeling, and frequency band reconstruction to restore the audio quality of MP3-compressed music. It divides the spectrogram into sub-bands, extracts gain-shape representations, and models both sub-band and temporal information for high-quality audio recovery. Trained with a Generative Adversarial Network (GAN), Apollo outperforms existing SR-GAN models on the MUSDB18-HQ and MoisesDB datasets, excelling in complex multi-instrument and vocal scenarios, while maintaining efficiency. - [2024.09.10] Apollo is now available on ArXiv and Demo. - [2024.09.106] Apollo checkpoints and pre-trained models are available for download. Apollo is trained on the MUSDB18-HQ and MoisesDB datasets. To download the datasets, run the following commands: During data preprocessing, we drew inspiration from music separation techniques and implemented the following steps: 1. Source Activity Detection (SAD): We used a Source Activity Detector (SAD) to remove silent regions from the audio tracks, retaining only the significant portions for training. 2. Data Augmentation: We performed real-time data augmentation by mixing tracks from different songs. For each mix, we randomly selected between 1 and 8 stems from the 11 available tracks, extracting 3-second clips from each selected stem. These clips were scaled in energy by a random factor within the range of [-10, 10] dB relative to their original levels. The selected clips were then summed together to create simulated mixed music. 3. Simulating Dynamic Bitrate Compression: We simulated various bitrate scenarios by applying MP3 codecs with bitrates of [24000, 32000, 48000, 64000, 96000, 128000]. 4. Rescaling: To ensure consistency across all samples, we rescaled both the target and the encoded audio based on their maximum absolute values. 5. Saving as HDF5: After preprocessing, all data (including the source stems, mixed tracks, and compressed audio) was saved in HDF5 format, making it easy to load for training and evaluation purposes. 🚀 Training To train the Apollo model, run the following command: 🎨 Evaluation To evaluate the Apollo model, run the following command: Here, you can include a brief overview of the performance metrics or results that Apollo achieves using different bitrates Different methods' SDR/SI-SNR/VISQOL scores for various types of music, as well as the number of model parameters and GPU inference time. For the GPU inference time test, a music signal with a sampling rate of 44.1 kHz and a length of 1 second was used. This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License . Apollo is developed by the Look2Hear at Tsinghua University. If you use Apollo in your research or project, please cite the following paper: For any questions or feedback regarding Apollo, feel free to reach out to us via email: `[email protected]`