Transcription Modes

A mode controls how HyperWhisper transcribes and formats your voice. Each mode stores one transcription engine, one language, the post-processing settings, and the formatting options. You can change the mode at any time to match your task.

​

Switching Modes

Press the Change Mode shortcut from any app to go through your modes. After each change, a notification shows the name of the active mode. You can also click the HyperWhisper icon and select a mode from the list.

​

Built-in Presets

When you create a mode, you select a preset. The preset sets the AI post-processing instructions for that mode. Both platforms give seven presets:

​

Transcription Source

Each mode has a transcription source: On Device, HyperWhisper Cloud, or Your Provider (bring your own API key).

​

On Device (local)

For fully offline transcription, download a Whisper, Parakeet, Nemotron, or Qwen3 ASR model. On macOS 26 and later, you can also use the Apple Speech Analyzer models. A large model is more accurate, but it uses more disk space. Whisper models with an .en suffix support English only. Parakeet v2 also supports English only, and Parakeet v3 supports many languages. You find and download models in Model Library. For the full list of on-device models and the platforms that support them, see Models.

​

HyperWhisper Cloud — Provider and Model

With HyperWhisper Cloud, you select the speech-to-text provider that processes your audio. You select the provider for each mode. Thus you can use a fast provider for messages, and a more accurate provider for meeting notes. Two rows control this. The Provider row names the company. The Model row names the model. The Model row is always a dropdown, also for a company that gives only one model. The Provider row shows ten companies, in alphabetical order: The Provider row shows (Recommended) on the ElevenLabs row only. ElevenLabs Scribe v2 is the recommended engine. The Model row shows no (Recommended) tag. Below the two rows, HyperWhisper shows the approximate credit cost for each minute of the model that you selected. A badge also appears there when the model is a Preview model, or when the model does not support custom vocabulary. Today every HyperWhisper Cloud model supports custom vocabulary, thus only the Preview badge appears. For your own provider (BYOK), some models still have no vocabulary support, and the mode editor shows a notice for them.

​

Your Provider (BYOK)

Select a cloud provider. Then enter your own API key in Model Library → API Keys. The model list changes with the provider that you select. Here also, the Model row is always a dropdown. A provider with one model shows that one model in the dropdown.

​

Language

You set the transcription language for each mode. Keep the value Auto if you want the engine to detect the language. Some local models and some cloud engines support English only. The app marks each combination that does not work. Tip: If you dictate in more than one language, create a mode for each language. This is more reliable than auto-detection for each recording.

​

AI Post-Processing

After transcription, HyperWhisper can send the raw text to an LLM. The LLM corrects the text, changes the format, and obeys the instructions of the preset. There are three options:

• Off — The raw transcription output, with no AI pass. This option is the fastest, and it does not change the format.
• Cloud — The AI pass runs on HyperWhisper Cloud, or with your own API key (OpenAI, Anthropic, Gemini, Groq, Grok, Cerebras, Mistral).
• Local — A model on your device processes the text. An internet connection is not necessary.

​

Raw Transcription (Post-Processing Off)

Set AI Post-Processing to Off to get the raw transcript from the STT engine. HyperWhisper makes no other changes. This is the fastest path. Use it when you want the exact words, or when you paste the text into a tool that applies its own format.

​

Cloud Post-Processing Engine & Model

If post-processing is Cloud and you use HyperWhisper Cloud, you can select the LLM engine and the model for this pass. This engine is not the transcription engine. You select the engine and the model for each mode. You can select Cerebras, Groq, Anthropic (Claude), Grok (xAI), OpenAI, Google Gemini, or Mistral. Each engine gives one or more models in a second picker. On both macOS and Windows, you can also send post-processing to a self-hosted or OpenAI-compatible custom endpoint instead of a built-in engine. See Custom Endpoints.

​

Additional Formatting Options

These options apply in addition to the preset instructions:

• Punctuation — Adds commas, periods, and other basic punctuation.
• Capitalization — Applies sentence case automatically.
• Profanity filter — Removes profanity from the transcript. The AI does not mask the words with symbols. If a sentence no longer makes sense without the word, the AI rewords that part slightly.
• Remove Trailing Period — Deletes the period at the end of a transcription. Use this option for short phrases that you paste into search fields, chat inputs, or command lines, where a final period looks wrong.

​

English Spelling Variants

If your language is English and post-processing is on, you can select a spelling variant. The AI then uses the same spelling rules for all text:

• American — for example, color, realize, organization, center, traveled
• British — for example, colour, realise, organisation, centre, travelled
• Australian — for example, colour, realise, organisation, centre, travelled
• Canadian — for example, colour, realize, organization, centre, travelled. Canadian English keeps the British -our and -re endings, but it uses the American -ize and -ization endings.

A new mode starts on the variant of your system region. A machine set to Canada starts on Canadian. A machine set to Australia starts on Australian. A machine set to the United Kingdom, Ireland, India, Singapore, South Africa, New Zealand, or another region that writes British English starts on British. All other regions start on American. An unknown region also starts on American. The region gives the first value only. HyperWhisper does not change the variant of a mode that already holds one. Thus a mode that you edited, and a mode that you restored from a backup, keep the value that they have. The variant also sets the date format that the AI writes: HyperWhisper stores the variant for each mode. Thus a mode for UK client emails can use British, and your daily dictation mode can stay on American.

​

Streaming Transcription

Streaming transcription shows the words while you speak. It does not wait until you stop the recording. It uses its own shortcut, not the standard toggle. The streaming settings are global on both platforms. Every streaming session uses them, and you cannot set them for each mode. For the full provider, language, and shortcut options, see Streaming Transcription.

​

Screen OCR

If Screen OCR is on for a mode, HyperWhisper reads the visible text on your screen when the recording starts. It sends this text to the AI as context. The AI then spells proper nouns, identifiers, and technical terms correctly. This function helps most when you dictate about the content on your screen. To use this function, edit a mode and turn on Screen OCR. At the first use, the app asks for the Screen Recording permission. To give this permission, see Permissions. Screen OCR is a setting for each mode. Turn it on only for the modes where the screen gives useful context (for example, a Code mode or a Meeting mode). Keep it off for casual dictation.

​

User-Supplied System Prompt

Every mode has an optional System Prompt field (up to 2,000 characters). HyperWhisper puts this text into the post-processing instructions, before the preset template. Use the field to add context that the preset does not cover. This context can include your industry, your preferred terminology, and your permanent rules for the output. Examples:

• “Use metric units. Spell out numbers below 10.”
• “This is for internal tech docs. Prefer ‘we’ over ‘I’.”
• “Always capitalize ‘HyperWhisper’.”

The system prompt applies to every transcription in that mode. It does not replace the preset instructions. It is an addition to them.

​

Gemini Custom Prompt

If the transcription source is Your Provider and the provider is Gemini, the mode gets a Gemini Custom Prompt field (up to 2,000 characters). This field sends your instructions directly to the transcription step of Gemini, before post-processing starts. The prompt controls how Gemini reads the audio. It does not control the format of the output. This field needs your own Gemini API key. The field does not appear for the Gemini models in the HyperWhisper Cloud Model row. This field is not the User-Supplied System Prompt. The system prompt applies to post-processing.

​

Cloud Transcription Domain

For HyperWhisper Cloud transcription, you can set a Transcription Domain on a mode. The audio then goes to a model for that domain. At this time, the only supported value is medical. This value selects a backend for clinical vocabulary and note formats. The Medical control appears only when the provider is AssemblyAI. AssemblyAI is the only provider that uses a domain for this. For Deepgram, medical is a model: select Nova 3 Medical or Nova 2 Medical in the Model row instead. You set this value for each mode, and HyperWhisper sends it to the cloud backend as a hint. For general-purpose transcription, keep the field empty.

​

Creating and Editing Modes

​

Tips

• One mode per context — Create a mode for each workflow (client emails, meeting notes, code, casual messages). This is better than a change of the settings in one mode. The Change Mode shortcut makes each change fast.
• Per-language modes — If you dictate in more than one language, a mode for each language is more reliable than auto-detection.
• Start with Hyper — The Hyper preset is sufficient for most general dictation. If you see the same gap many times, add a system prompt or select a different preset.
• Raw output for other tools — If another tool applies its own format (IDE, terminal, form field with validation), set post-processing to Off.