Recommended specs and speed

SAYRIL runs on Windows 10 and 11 and Linux (64-bit). A graphics card is not required: without one, recognition runs on the processor.

System

Windows 10 or 11, 64-bit (x64); Windows 10 version 1809 or later. No administrator rights needed

System

Linux, 64-bit (x86-64): Ubuntu 22.04 or later, Debian 12 or later, Linux Mint 21 or later, Astra Linux 1.8; Astra Linux SE 1.7 has a separate build. DEB package or AppImage

Microphone

A microphone the system detects. Recognition quality depends on the microphone and background noise; it has not been measured with different microphones. Meeting recording also captures computer audio (on Linux, only the microphone)

Activation

The program computes the device code from five hardware and system details. If it reads fewer than three of them, the program cannot be activated on this computer. In a virtual machine without hardware information, support provides activation on request

Graphics card

Optional. Acceleration: NVIDIA via CUDA, and AMD, Intel and NVIDIA via Vulkan. Only the regular driver is needed; for CUDA, one that supports CUDA 12. Recommended free video memory, as measured on Windows: 2.0 GiB for recognition via CUDA and 0.9 GiB via Vulkan, 1.4 GiB for translation, 2.3 GiB for text editing by the model. On Linux, running on graphics cards has not been tested yet; running on the processor has been tested

Processor

The program as a whole needs an x86-64 processor with SSE4.2 and POPCNT (the x86-64-v2 level). Translation, text editing by the model and recognition via Vulkan also need AVX2, FMA, F16C and BMI2. Intel has made such processors since 2013 and AMD since 2015, but check support for your processor model. Dictation on a processor without AVX2 has not been tested

RAM

The recommended amount for each feature is SAYRIL's measured peak plus the memory Windows itself needs (2 GB for Windows 10 and 4 GB for Windows 11, according to Microsoft). On Windows 10 and Windows 11 respectively: dictation 3.8 and 6.6 GiB; translation 8.0 and 9.7 GiB; text editing by the model 9.3 and 11.3 GiB; dictation with translation and editing together 13.0 and 15.0 GiB; transcribing a file up to 3 hours long and recording a meeting with speaker separation 4.4 and 7.4 GiB. The translation and editing figures are for a computer with integrated graphics, which takes its memory from RAM. With less than 6 GiB of RAM, the program switches to a low-memory mode on its own: recognition runs on the processor and is unloaded before translation and editing. Memory use on Linux has not been measured. You can try SAYRIL on your own computer during the free 72-hour trial

Disk space

On Windows, you need 14 GiB of free space during installation: that is how much the installer, the archive and the unpacked models take up while unpacking is in progress. Once installed, SAYRIL and its models folder take up 6.95 GiB, or 7.70 GiB if you select “Graphics card” during installation (this adds the CUDA libraries for NVIDIA). Do not delete the models folder: SAYRIL loads the models from it. On Linux, the installed DEB package takes up 1.25 GiB, the AppImage file 0.70 GiB, and the models folder another 6.65 GiB. While the models are being installed, their archive needs another 6.65 GiB

Internet

Needed to download the files and get the license file. Not needed after activation

1 GiB = 1,073,741,824 bytes. If you select an existing models folder during installation, the models are not copied a second time.

• Recognition: about 0.3 s per 8–9 s phrase on a graphics card; 0.3–2 s on the processor (Russian and European languages are faster).
• Translating one phrase: about 0.2 s via Vulkan and about 0.9 s on the processor.
• Russian on the processor: 4.8% word errors. 775 phrases from FLEURS-ru, punctuation not scored.

Your computer will give different numbers. The program has a “Tests and performance” page: you start the measurement manually, and the report stays with you.

• On Wayland, programs cannot see key presses in other windows: dictation keys are assigned in the system settings
• On Wayland, text is inserted automatically if wtype or ydotool is installed; otherwise it stays on the clipboard
• On Linux, meeting recording captures only the microphone, without computer audio
• On Linux, running on graphics cards has not been tested yet; running on the processor has been tested
• On Linux, files other than WAV are read by a system decoder, if one is installed