Configuration

Every option has a working default, so configuration is optional. To change one, override the row in $DSH_HOME/profiles/web/cordis.patch.yml — a patch replaces a row's entire config, so restate every key you still want:

- id: voice-input
  name: dsh-voice-input
  config:
    model: small.en
    language: en
    computeType: int8
    timeoutMs: 300000

Key	Default	Meaning
`model`	`base.en`	Whisper model size or name
`language`	`en`	Spoken language, or `auto` to detect
`computeType`	`int8`	CTranslate2 compute type
`timeoutMs`	`300000`	Deadline for one transcription
`polish`	`conservative`	Clean the transcript with a model; `off` returns the raw text
`polishTimeoutMs`	`15000`	Deadline for the cleanup call alone
`polishProvider` / `polishModel`	current selection	Pin cleanup to a specific route
`pythonPath`	package `.venv`	Interpreter with `faster-whisper`
`scriptPath`	bundled CLI	The transcription script
`cacheDir`	`$DSH_HOME/cache/voice-models`	Model weights cache

The cleanup uses the model the deployment already selects, so it needs no second credential and no extra configuration. It is a normal ctx.llm.stream call — the same route the agent uses — and it runs with temperature: 0 under its own deadline.

Choosing a model

Model	Size	Notes
`tiny.en`	~75 MB	Fastest, least accurate
`base.en`	~150 MB	Default; roughly 1 s of CPU per 5 s of speech
`small.en`	~500 MB	Noticeably better, about 3× the compute

Drop the .en suffix for multilingual models (base, small) and pair them with language: auto. A model that is not cached downloads on first use and needs network access once.

If huggingface.co is blocked, set DSH_VOICE_HF_ENDPOINT to a mirror such as https://hf-mirror.com before starting DSH.

dsh-voice-input

Other repositories with this package name

Install

README

Configuration

Choosing a model