Repository navigation
Adds Linux ARM64 with automatic Vulkan Engine selection and no CUDA requirement.
The new ARM64 Debian package depends on openresearchtools-engine (>= 1.17).
APT selects and downloads the ARM64 Vulkan engine automatically from the Open Research Tools repository.
For manual installation, download engine-arm64.deb
and install it together with the new app package using sudo apt install ./engine-arm64.deb ./transcribe-offline_2.1.1_arm64.deb.
Windows x64, macOS ARM64, and Linux AMD64 binaries are copied byte-for-byte from release 2.1.
Their embedded versions and behavior remain unchanged. Only the Linux ARM64 application is newly built.
Downloads
- transcribe-offline_2.1.1_arm64.deb
- Transcribe-Offline-windows-x64.exe
- transcribe-offline-macos-arm64.dmg
- transcribe-offline_2.1.0_amd64.deb
- SHA256SUMS.txt
Optional Whisper compatibility setting
In Settings → Transcription Settings, enable Mobile GPU compatibility (Whisper) to set --whisper-beam-size 1 --whisper-best-of 1. Some mobile GPUs do not work with larger decoding values. The setting is optional, off by default, saved between launches, and applies with or without speaker diarization. It uses existing Engine CLI options; engine v1.17 does not need rebuilding.
Validation
Native ARM64 build and 52 automated tests passed. On Ubuntu ARM64 with Adreno X1-45, short WAV and MP3 clips using Whisper tiny English matched the reference exactly with both default and single-candidate decoding. The compatibility option also passed with Sortformer diarization, producing a speaker-labelled transcript. Final inference tests used a 2 GiB memory limit with swap disabled. Desktop startup, Vulkan detection, and migration of saved CUDA settings were also checked.