Local models #

On-device models for titles, judgments, translation, speech, dictation. Never downloaded automatically.

eukhe tiny-models                       # list models and state
eukhe tiny-models download lfm2.5-230m  # download one, or all
eukhe translate de hello                # translate to German
eukhe say "hello" [--out a.wav]         # speak, or write a WAV
eukhe transcribe a.wav                  # WAV to text

• Text: lfm2.5-230m, lfm2.5-350m. Speech to text: whisper-base, whisper-small. Text to speech: kokoro.
• Stored in ~/.eukhe/models/tiny/; download checks size and sha256 per file.
• ONNX Runtime 1.24+. Nix sets ORT_DYLIB_PATH to nixpkgs' library; yours wins.

Roles #

A role uses the first available of its model, then its fallbacks (local: installed; provider: the session's model). None: the task fails.

Role	Task	Default order
tiny	Session titles	local/lfm2.5-230m, local/lfm2.5-350m, the session's model
judge	Typed judgments	As tiny
translate	Translation	As tiny
speech	Text to speech	local/kokoro
dictation	Speech to text	local/whisper-base, local/whisper-small

Key	Default	Effect
modelRoles.<role>	none	Role model: local/<id> or provider/model. /model hub roles titles, judge, translate, speech, dictation set it globally (settings.md); clearing there falls back to the default order
retry.fallbackChains.<role>	default order	Models after modelRoles.<role>; [] = only modelRoles.<role>
tts.localVoice	af_heart	eukhe say voice; --voice overrides
stt.enabled	false	Dictation into the editor via dictation
providers.tinyRuntimeLibrary	none	ONNX Runtime library path; none = ORT_DYLIB_PATH (global only)

Dictation #

set stt.enabled true                 # config.cfg
bind <key> "action app.stt.pushToTalk" # binds.cfg: hold to talk; or action app.stt.toggle; neither bound by default

Recorder: first on PATH of pw-record, parec, arecord (macOS: sox, then ffmpeg).