BetterWispr Review: Local Mac Dictation

A shortcut, then a finished transcript

By default, you hold the shortcut while speaking and release it when you're done. You can switch to a press-to-toggle mode or assign a different shortcut using Command, Option, Control, or an F key.

One thing to know if you're used to live captions: BetterWispr starts transcription after you stop recording. It's built around a short record-and-insert loop, not words appearing in the target app while you speak. If you want to monitor the transcript as you go, that difference may affect whether it fits your workflow.

If insertion doesn't work, BetterWispr copies the text to the clipboard so you can paste it manually. Some apps - and secure fields - may block automatic insertion. In those cases, the app includes a dashboard where you can retrieve the transcript and insert it yourself.

Cleanup and writing style are separate controls

Dictation output doesn't always come out ready to send. People repeat themselves, use fillers, restart sentences, and leave punctuation implicit. BetterWispr includes cleanup levels to help bridge that gap:

• Light removes common English filler words, repeated words, and false starts.
• Medium also edits for clarity using the notes model selected in settings.
• None keeps every word, which helps when exact phrasing matters.

You can also pick a tone depending on the app. For instance, a casual chat style can drop the final period, and a very casual style can drop capitalization too. A formal style leaves your wording alone. Styles are applied to English, so you shouldn't assume they'll behave the same way for every language supported by the underlying speech model.

Keeping cleanup and style separate is actually helpful. The transcription model handles what it heard; cleanup controls how much to revise; style controls how the final text is presented. Same input, different knobs. That makes it easier to use a conservative setup for important messages and a looser one for informal notes.

Vocabulary improves repeated names and terms

Names, product terms, and technical vocabulary are frequent sources of transcription errors. BetterWispr lets you add words as hints and replacements in Vocabulary. You can also correct a word in History or right after the text is pasted. The app can learn that correction and reuse your preferred spelling next time, and learning can be turned off in Settings.

This is a practical feedback loop - not a promise that speech recognition will nail every name on the first try. If you dictate the same client names or technical terms regularly, a small vocabulary list can cut down on repeated fixes. If your terminology changes a lot, you may prefer to review what it has learned and keep the list up to date.

Choose a built-in model or connect a provider

BetterWispr includes nine built-in models from Apple speech, Parakeet, and Whisper. Language coverage varies by model. Parakeet TDT v3 supports 25 European languages, there's a separate Parakeet model for Japanese, and Whisper provides broader multilingual recognition. Apple speech depends on your locale and what system assets are installed.

You can also choose your own Sarvam AI, Smallest AI, or an OpenAI-compatible API. That changes the privacy and infrastructure picture. A local model processes speech on the Mac; a remote API sends data to the provider you configure. The app's "local-first" description shouldn't be treated as a guarantee that every configuration will keep everything local.

The same idea applies to writing assistance. Apple Intelligence or Ollama can handle Medium cleanup and generate meeting summaries on your Mac, as long as your setup supports it (OS version, hardware, and model configuration). Claude Code, Codex, and other API connections are optional, but using them sends the relevant content to those providers.

If you want a simple rule: decide what data path you're comfortable with before you pick a model. For personal notes or sensitive work, make sure both the speech and cleanup models are local. If you intentionally choose a provider, read that provider's data handling terms and compare them with your organization's policies.

History gives you a way to review and restore

BetterWispr saves dictations on your Mac, including both the raw text and the cleaned-up text. The Use Original option restores what you said before cleanup. History can also be disabled.

Keeping the raw text is useful. If cleanup changes a sentence more than you expected, you can compare it with the original instead of trying to reconstruct your wording from memory. It also means you should think about whether you want history at all. Local storage isn't the same as "no storage," and people sharing a Mac - or handling sensitive information - may want to turn it off.

The app also tracks things like word counts, pace, words removed by cleanup, the apps you dictation into, and daily streaks. Those numbers are calculated from history on your Mac. They can help you understand your usage, but they aren't required for dictation. If you don't want a record of your activity, check the history settings instead of assuming analytics won't exist just because processing is local.

Meeting notes add a different workflow

Notetaker records your microphone as Me. On macOS 14.2 or later, it can also capture audio from other apps as Them. You can write your own notes alongside the transcript. Afterward, Apple Intelligence on supported Macs running macOS 26 - or a local Ollama model - can generate a summary, decisions, and action items.

Even without a notes model, the transcript and your notes are still saved. Optional connections like Claude Code, Codex, and OpenAI-compatible endpoints can also generate notes, but in those cases the transcript gets sent to the provider you chose.

This is handy when you want both a transcript and your own observations in one place, with an optional summary after the meeting. It isn't a substitute for checking what was said. Speaker labels and action items can be wrong, and sensitive meetings may have consent, workplace, or legal requirements. Make sure recording is appropriate before capturing anyone else's audio.

Install, grant permissions, and select a model

Install the app and follow its welcome guide. Initial setup includes permissions and installing your first local speech model. BetterWispr may request:

• Microphone permission to record your voice.
• Accessibility permission to insert text into another app.
• Speech Recognition permission when using Apple speech.

These permissions serve different purposes. If insertion fails, check the target app and Accessibility permission first. If you don't grant permission for insertion, you may still be able to retrieve the transcript from the dashboard and paste it manually.

Before you rely on it for important work, test the model with your own voice, accent, vocabulary, and typical recording environment. A model's language list is just a starting point. Noise, specialist terms, and how fast you speak all affect real accuracy.

Set the interaction and cleanup to match the task

Place the cursor in a text field and hold Option + Space while speaking. Release the shortcut to finish and insert the transcript. If holding the shortcut feels awkward, switch to press-to-toggle or pick a different shortcut in Settings.

Then tune output in stages:

1. Start with None if you want a clear baseline for what the model recognizes.
2. Try Light for everyday messages where removing English fillers and false starts helps.
3. Use Medium only when you want additional clarity edits and you've selected a notes model you trust.
4. Set app-specific tones for cases where punctuation and capitalization should differ.
5. Add recurring names and terms to Vocabulary, then check History when a correction matters.

This sequence helps you separate "recognition was wrong" from "editing preferences changed the output." If a name is wrong in the raw transcript, adjust the model or Vocabulary. If the raw text is fine but the final sentence changes too much, reduce cleanup. If everything is accurate but it just doesn't look right, change the tone for that app.

Treat model updates and app updates differently

Speech models live on your Mac and use local storage. BetterWispr updates for the app come through GitHub Releases via Sparkle, and they install only when the EdDSA signature matches. The project is licensed under Apache 2.0, and you can inspect or build the source with make dev.

For most people, the practical point is that the app and the speech models are separate parts of the setup. Updating the app doesn't remove the need to choose and install a model that works for your needs. If you're evaluating an open-source tool for a team, review the source and release process as part of your normal software checks, instead of treating "open source" as an all-in-one security review.