AI enhancement

Enhancement takes the raw transcript and runs it through a language model —
fixing punctuation, structure, and the words a speech model reliably mishears.

Whether it runs is decided by one setting: Output, at the top of the Enhancement
page. Instant never uses AI; Instant + Refine and Enhanced do.
Zerm dictates perfectly well on Instant.

Each dictation is enhanced with exactly the prompt, provider, model, and context it
started with. Editing settings while a dictation is being processed does not change it;
the recorder shortcuts are how you change the one in
progress.

Providers

Enhancement is provider-agnostic. Whatever you pick, the transcript and the active
prompt are what get sent, and nothing else leaves your Mac unless you switch context
options on.

On-Device. A local model runs through llama.cpp with no network access at all:
Gemma 4 E2B by default, or Qwen3 4B Instruct as a smaller option. It is the only
provider that keeps enhancement entirely offline.

Ollama. Points at a local Ollama server, so the model still runs on your own
hardware — useful if you already keep models there.

Local CLI. Shells out to a command-line tool you nominate. For anything you host
yourself that has no HTTP API worth wiring up.

Cloud. Cerebras, Groq, Gemini, Anthropic, OpenAI, OpenRouter, and Mistral are
configured with an API key. A Custom option accepts any OpenAI-compatible chat
completions endpoint.

No provider is contacted until you add a key and select it. A key is verified when you
add it, and is not saved if verification fails. Providers that only transcribe speech
are not offered here.

Prompts

A prompt is the instruction the model gets alongside your transcript. Zerm ships with
Default, Assistant, Coding, and Chat, and you can add your own from
templates or from scratch.

The active prompt is what runs. Switch it from the recorder with ⌘1 through ⌘0 — see
enhancement shortcuts — or pin a prompt to a
Power Mode so it changes with the app you are in.

A provider and model per prompt

A prompt can use its own AI Provider and AI Model instead of the global ones. A
quick cleanup prompt can stay on the on-device model while a long rewrite uses a larger
cloud model. When both set one, a Power Mode's choice wins over the prompt's.

May change language or length

Enhancement is normally not allowed to translate or to balloon your text; see
the language section below. A prompt whose job is exactly that — translating,
expanding notes into an email — can turn on May change language or length, and those
checks are skipped for it. The built-in Assistant prompt has it on.

Trigger words

The prompt editor has a Trigger Words field. Trigger-word switching is currently
turned off, so starting a dictation with a trigger word does not change the prompt.

Context

By default the model sees the transcript and nothing else. Two optional context sources
can be added in Enhancement settings; each is its own switch.

Clipboard Context. The clipboard is read when the recording starts, not when the
model runs. Useful when you are dictating a reply to something you just copied. Used by
both Instant + Refine and Enhanced.

Screen Context. On-screen text is captured and included. Needs Screen Recording
permission. Used only in Enhanced: Instant + Refine has already pasted by the time
the screen could be read.

The details of what is captured, when, and what is kept are on the
contextual awareness page. Your
dictionary vocabulary is also passed along as spelling guidance
whenever enhancement runs.

Language

Enhancement cleans the transcript. It does not translate it.

Mixed-language dictation stays mixed. Hebrew stays Hebrew, English stays English,
Russian stays Russian. Zerm discards a rewrite and keeps your original text when the
model:

• returns a different writing system than you spoke,
• grows the text far beyond what you said, or
• starts talking about itself instead of rewriting.

Prompts with May change language or length are exempt.

If you always speak one language in an app, pin that language on the
Power Mode rather than leaving Auto. Pinning Hebrew when you are
about to speak English will make the speech model hear everything as Hebrew.

Skipping short transcriptions

"Yes", "thanks", and "on my way" do not benefit from a language model, and sending them
adds latency for nothing. Skip short transcriptions is on by default and bypasses
enhancement whenever the transcript is at or under Minimum words, which defaults to
three and is adjustable from one to fifteen.

Timeouts and retries

Timeout duration bounds how long Enhanced output waits for the model, from 3 to 60
seconds. The default is 15. Instant + Refine runs after the paste has already happened
and uses its own, much shorter limit.

On timeout decides what happens when the time runs out: Retry (the default), up to
three attempts, or Fail immediately. Either way you end up with the raw transcript
rather than nothing — a slow model costs you the improvement, never the words.

Network errors, server errors, and rate limits are retried with a short backoff.

When enhancement does not run

Zerm tells you when it cannot enhance, instead of silently pasting the raw text. A
notification explains what happened, and your original text is always kept:

• Skipped — the provider is not set up: a model is not downloaded, an API key or
  model is missing, or a server address is wrong. The notification links straight to
  Enhancement settings, and appears once per launch.
• Failed — the provider could not be reached, returned nothing, hit a rate limit, or
  had a server error.
• Timed out — the timeout ran out.
• Discarded — the model changed the language or the wording too much.

Enhancement is skipped without a notification when:

• The output is Instant, globally or in the active Power Mode.
• The transcript is at or under the short-transcription threshold.