How to change the Whisper model¶

Available models¶

Model	Size	VRAM (int8)	Speed	Notes
`medium`	1.5 GB	~700 MB	~0.5s	Multilingual, recommended default
`medium.en`	1.5 GB	~700 MB	~0.5s	English-optimized, slightly better for English
`large-v3-turbo`	1.6 GB	~2.3 GB (float16)	~0.3s	Fastest but higher VRAM
`small`	466 MB	~300 MB	~0.8s	Low VRAM, lower accuracy

Edit your systemd service:

systemctl --user edit mirach

Add:

[Service]
Environment=MIRACH_WHISPER_MODEL=medium.en
Environment=MIRACH_WHISPER_COMPUTE=int8

Restart:

systemctl --user restart mirach

The model is downloaded automatically on first use.

Compute	Device	VRAM	Accuracy
`int8`	cuda or cpu	Low	Good for speech
`float16`	cuda only	High	Maximum accuracy
`int8_float16`	cuda only	Medium	Hybrid approach

For most users, int8 on cuda is the best balance.