Everything is an environment variable prefixed SCRIBE_, read once at start by scribe/config.py. Put them in your shell profile or a file you source before scribe serve.
tiny.enbase.ensmall.enmedium.enlarge-v3, or a path to a CTranslate2 model folder
SCRIBE_STT_DEVICE
auto
auto picks cuda if available, else cpu
SCRIBE_STT_COMPUTE
auto
float16 on GPU, int8 on CPU. int8_float16 saves VRAM
SCRIBE_STT_LANGUAGE
en
Empty string for auto-detect (multilingual models only)
SCRIBE_STT_BEAM
5
Beam size; 1 is faster and slightly worse
SCRIBE_VOSK_MODEL
models/vosk-model-small-en-us-0.15
Folder of an extracted Vosk model
The scribe also sets ORT_DISABLE_TELEMETRY=1 for you. That stops the onnxruntime library, which Whisper uses to skip silence, from reporting usage to Microsoft. Set it to 0 before starting if you want the library's own default back. All open source, no cloud has the measurement.