> ## Documentation Index
> Fetch the complete documentation index at: https://docs.openwhispr.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Getting better accuracy

> Practical things that improve transcription quality — the engine you use, your microphone, your language setting, and teaching it your vocabulary.

If transcription is close but not right, there are four levers worth pulling, in
roughly the order they pay off.

## 1. Name your language

Auto-detect is the default, and it's the weakest setting for accuracy because
the model spends part of its effort deciding what language you're in. If you
nearly always dictate in one language, set it explicitly under **Settings** →
**Preferences** under **App**, in the **Language** section.

This is the single cheapest improvement for most people.

## 2. Teach it your words

Names, jargon, product names and acronyms are what transcription gets wrong most
often, and no amount of model tuning fixes a word the model has never seen.

The **custom dictionary** exists for this — add the terms you use and they'll be
recognised. See [custom dictionary](/help/customise/custom-dictionary).

<Tip>
  If you've named your voice agent, its name is added to the dictionary
  automatically so it's recognised reliably when you address it.
</Tip>

## 3. Check your microphone

The built-in microphone on a laptop is usually the weakest link, especially in a
room with any background noise. A headset — even a cheap one — is a bigger
improvement than changing models.

Pick which microphone OpenWhispr uses under **Settings** → **Preferences** under
**App**. If it isn't picking up at all, see [my microphone isn't
working](/help/fix/microphone-not-working).

Speaking at a normal pace helps more than speaking slowly and clearly. These
models are trained on natural speech, and over-enunciating is further from what
they expect, not closer.

## 4. Choose a better engine

Transcription runs on one of several engines, and they aren't equally accurate.
Set yours under **Settings** → **Speech-to-Text** under **AI Models**, on the
tab for the mode you're changing — **Dictation**, **Note Recording**, or
**Audio Upload**.

| Engine               | Trade-off                                                              |
| -------------------- | ---------------------------------------------------------------------- |
| **OpenWhispr Cloud** | Most accurate, needs a connection                                      |
| **Whisper (local)**  | Private and offline; accuracy scales with model size and your hardware |
| **Parakeet (local)** | Faster than Whisper locally, fewer languages                           |
| **AssemblyAI**       | Your own key, 7 languages                                              |

If you're running a local model and accuracy is poor, a larger model is usually
the answer — see [local models](/guides/local-models).

## What cleanup does and doesn't fix

Dictation cleanup tidies grammar and punctuation, but it can't recover a word
that was misheard — it only sees the transcript, not your audio. Persistent
wrong words are a transcription problem, not a cleanup one.

<Snippet file="still-need-help.mdx" />

## Related

* [Custom dictionary](/help/customise/custom-dictionary)
* [Dictating in other languages](/help/dictation/languages)
* [The words come out wrong](/help/fix/wrong-words-or-language)
* [Local models](/guides/local-models)