Most Mac dictation apps were built for marketing emails. Developers have specific needs: technical vocabulary, library names, dictating into Cursor and chat, prose around code rather than the code itself. Three apps that don't fall over on developer workflows.
Verified May 2026 · written by the Murmur team
Built-in dictation models are trained on consumer speech: "let's meet at noon," "send the kids to school," "what's the weather like." They don't expect "deploy the staging Kubernetes cluster," "kafka consumer group offset," or "patch the OAuth scope on the JWT." The result is mush.
Whisper, the model behind Murmur, MacWhisper, VoiceInk, and SuperWhisper, was trained on a much broader corpus including technical podcasts and conference talks. It handles the long tail of programming vocabulary noticeably better. Not perfect (it still hears "Vue" as "view" sometimes) but enough that you stop wincing.
Dictation is bad at writing code. The accuracy problem isn't the issue: the syntax problem is. Saying "open paren x equals one comma y equals two close paren" is slower than typing. And speech grammar collapses on nested structures.
What dictation is great at: the prose around code.
If you live in Cursor, the workflow is: open the chat panel, double-tap right Option, mumble what you want. Murmur takes a screenshot of the current Cursor window (so Claude sees the file, the line you're on, the surrounding code), transcribes your speech, and pastes a polished prompt that includes context.
The friction this removes is small per-turn but stacks up. You don't have to type "make this function async" with the file path and line number. You just look at the line and speak. The screen-context capture handles the rest.
The same pattern works in Claude Code (the CLI), Continue.dev, and any web-based AI chat. The screen capture means the AI sees what you're looking at, which is usually the answer to "what context do you need?"
Whisper transcription handles technical vocab. Screen-context AI mode pastes Cursor-aware prompts. Audio stays on-device, no proprietary code goes to a third-party cloud. Claude API spend is yours.
Read the source, build from source, run the LLM step on-device via Ollama. The "I want everything inspectable and no cloud anywhere" choice. Cheaper than Murmur if you can live without screen-context.
Best polish engine in the category. Per-app tone profiles mean Slack vs Linear vs your IDE all get appropriate output. Audio is uploaded to their cloud, which can be a non-starter for proprietary code or NDA work.
If you're an indie developer working on your own SaaS, the privacy stakes are personal: your code, your business. If you work at a company with an NDA, the stakes are professional: lots of CISOs would have feelings about your audio being transcribed in someone else's cloud.
The safer cluster is anything where audio stays local: Murmur, VoiceInk, MacWhisper, SuperWhisper, Apple Dictation. Within that cluster, the LLM polish step is a separate decision. With Murmur, the polished text goes to Claude via your key (Claude API traffic is not used for training by default). With VoiceInk + Ollama, even the LLM stays local.
Worth running your own check with Little Snitch. Watch what each app actually does on the wire.
Murmur ($29 lifetime) for Cursor + screen-context AI. VoiceInk ($25 lifetime) for open-source + local LLM. Wispr Flow ($12/mo) for the most polished output across every IDE. Apple Dictation struggles with technical vocabulary and is rarely the right pick.
Whisper-based apps handle technical terms, library names, and function names noticeably better than Apple Dictation. None of them write syntactically perfect code from speech. Use dictation for the prose around code (PR descriptions, Slack messages, git commit messages, comments) and let Cursor or your IDE handle the actual code.
Yes. Mac dictation pastes wherever the cursor is, so it works in Cursor's chat, VS Code's terminal, comments, anywhere. Murmur's screen-context AI is particularly strong for Cursor: double-tap, mumble what you want, Claude reads what's on screen and writes a polished prompt.
With Murmur Quick mode, no, fully local. With Murmur AI mode, transcribed text and a screenshot go to Claude via your key. With VoiceInk + Ollama, fully local. With Wispr Flow / Aqua Voice, audio is uploaded to their cloud.
Whisper handles uncommon technical names better than Apple Dictation out of the box. For project-specific terms (a custom library, a colleague's name, a code-name product), Murmur's AI mode prompt can include "the project is called X, the team is Y, Z" so Claude corrects mishears in polish. SuperWhisper has explicit custom-vocabulary lists.
Murmur turns rough speech into polished prompts for Cursor and Claude, with screen-context so the AI sees the file you're working on. Audio stays local; AI calls go through your key.