v3.21.0 - Update GPU Dockerfile to CUDA 11.8 and PyTorch to 2.4.0

Changed

• Update GPU Dockerfile to CUDA 11.8 and PyTorch to 2.4.0
• Update dependencies

v3.20.2 - Fix iOS support

Added

• Add missing dependency for building on AArch64

Changed

• Fix iOS support for exporting mixes (#1145) - credit to @microtherion for investigating
• Update dependencies

v3.20.1 - Fix first-time Spleeter model download error

Changed

• Fix regression where first-time Spleeter model download fails due to httpx not following redirects (#949) - thanks @microtherion for the fix!
• Add cleanup logic to remove leftover model files from failed download attempts
• Update dependencies

v3.20.0 - ARM/AArch64 Docker support

Added

• Add Docker support for AArch64 systems (including Apple Silicon)

Changed

• Update Docker Python version to 3.9
• Update Docker CUDA image version to 11.6.1
• Update dependencies

v3.19.0 - Add music player download button and yt-dlp options

Added

• Add 2 environment variables for configuring yt-dlp:
  • YOUTUBEDL_SOURCE_ADDR - Client-side IP address for yt-dlp to bind to. If you are facing 403 Forbidden errors, try setting this to 0.0.0.0 to force all connections through IPv4
  • YOUTUBEDL_VERBOSE - Set to 1 to enable verbose logging for yt-dlp
• Add download button on music player

Changed

• Update dependencies

v3.18.0 - Bug fixes and improvements

Added

• Add AWS_S3_REGION_NAME and AWS_S3_SIGNATURE_VERSION variables for S3 integration
• Add more descriptive error logging

Changed

• Fix bug where separation would fail immediately (#672)
• Allow multiple CERTBOT_DOMAINs and hardcode cert-name to 'spleeter-web' (#533)
• Update GPU Docker base image to cuda:11.2.2 from cuda:11.2.0 due to deprecation
• Update dependencies

v3.17.0 - Spleeter 5-stem model (with piano)

Added

• Support Spleeter's 5-stem model which separates piano in addition to other accompaniment

Changed

• Update separator form UI to use buttons instead of dropdown menus
• Improve form field validation
• Bring back 'iterations' param for X-UMX model
• Fix Dev Docker Compose config
• Fix bug where deleting dynamic mix may cause blank screen
• Update dependencies

Removed

• Remove frontend upload size check and rely on Nginx and backend instead

v3.16.0 - Lossless output formats and configurable upload limits

Added

• Support lossless output formats (WAV, FLAC)
• Add UPLOAD_FILE_SIZE_LIMIT (MB) and YOUTUBE_LENGTH_LIMIT (minutes) environment variables to customize upload limits

v3.15.0 - Demucs v4 and configurable segment sizes

Added

• Update Demucs to v4, which features new Hybrid Transformer models with average SDR of 9.0
  • More info
• Add ability to configure segment sizes for Demucs models
  • If Demucs separation fails for you due insufficient GPU memory, try setting the environment variable DEMUCS_SEGMENT_SIZE to a lower value like 10 (default is 40, which requires a around 7 GB of memory). You may have to experiment a bit to find the appropriate value. See this for more info.
  • Also try setting PYTORCH_NO_CUDA_MEMORY_CACHING=1 to disable caching

Changed

• Update Docker Compose definitions from version 3.4 to Compose Specification
  • You may have to update your Docker Compose version to the newer version

v3.14.0 - Raise minimum requirements and dependency updates

Changed

• Raise minimum Python version to 3.8 and NodeJS version to 16
• Update Docker images to use Python 3.8 and NodeJS 16 (bullseye)
• Update Spleeter, Demucs, Tensorflow, and other dependencies

Removed

• Remove some unused NPM dependencies