What voice prompting actually is
Voice prompting is a two-step pipeline: speech-to-text, then text-to-AI. You speak. A speech model (typically Whisper) turns your voice into text. That text is sent to an AI model — ChatGPT, Claude, Gemini, or anything else with a chat-completions endpoint — as the prompt. The AI's reply is what gets typed at your cursor.
The result is a workflow where your input is voice, the model's reply is text in the place you wanted it, and you never opened a chat window. The browser tab stays closed. The IDE stays focused. The Slack thread stays scrolled to where it was.
Why people switch to voice prompting
Three measurable wins appear in every workflow study we've seen on this:
- Throughput. Comfortable speech is 130–160 words per minute. Comfortable typing is 40–70. Even after editing time, voice-prompted AI runs roughly 2× more iterations per hour.
- Lower context-switch cost. Tab-switching to a chat UI breaks flow. Hotkey-and-speak doesn't.
- Better prompts. Spoken prompts tend to be more conversational and more complete. Typed prompts tend to be terse to save keystrokes — which gives the model less to work with.
Voice prompting ChatGPT
ChatGPT itself has a voice button on web and a voice mode on mobile. Both are great inside the ChatGPT app. They are not great when what you want is for the answer to land inside the document you're writing in Word, the comment you're leaving on a GitHub PR, or the email you're composing in Outlook.
The desktop voice-prompting setup that solves this looks like:
- Install an offline voice tool that supports global hotkey dictation and BYOK AI routing — AirTypes does this.
- Paste your OpenAI API key into the My Agent settings. Pick a model (GPT-4o or GPT-4 Turbo are the common picks).
- Place your cursor where you want the answer to appear.
- Hold the agent hotkey, speak your instruction, release.
- The text of your instruction is transcribed locally, sent to OpenAI through your key, and the model's reply is typed at your cursor.
Voice prompting Claude
Anthropic's Claude API exposes the same OpenAI-compatible chat completions shape that the rest of the industry has converged on. Two paths to voice-prompt Claude:
- Direct via Anthropic API key. Point your dictation tool's My Agent at the Anthropic endpoint with your key. Claude Opus, Claude Sonnet, and Claude Haiku are all available.
- Via OpenRouter. One key, hundreds of models, including the entire Claude family. Useful when you want to switch between Claude, GPT, and Gemini without juggling three providers.
Many teams default to Claude for long-form analysis and writing tasks because of its strong instruction-following on multi-paragraph prompts — exactly the kind of prompt voice tends to produce.
Voice prompting Gemini
Google's Gemini family is available either through the Gemini API directly or through OpenRouter. Google has been shipping OpenAI-compatibility endpoints alongside their native API specifically so tools like this don't need a custom integration.
Gemini's strengths in this workflow are large context windows (handy for "summarise this long doc" voice prompts) and competitive Flash-tier latency (handy when you want the response to feel conversational).
Voice prompting local models
For full privacy — including the AI inference step — point My Agent at a local Ollama or LM Studio server. The whole round-trip stays on the device:
- Microphone capture — local.
- Whisper transcription — local.
- Inference (Llama 3.x, Mistral, Qwen, Phi, DeepSeek) — local.
- Response typed at your cursor — local.
This is the deployment shape that legal, healthcare, and government teams use when no part of the workflow is permitted to leave the endpoint.
Profile patterns that actually work
The single biggest accuracy upgrade in voice prompting isn't a better speech model — it's giving the AI a system prompt that matches what you're trying to do. The fastest way to do this is via named profiles you trigger by saying their name first.
A short list of profile patterns that earn their keep on day one:
- Email — "Rewrite the user's words as a clear, professional email body. Return only the body."
- Reply — "Compose a short, polite reply to the email or message above the cursor based on what the user dictates."
- Summary — "Summarise the user's words into a concise paragraph. Preserve facts and numbers."
- Bullet — "Convert the user's words into a clean bulleted list with parallel structure."
- Code — "Treat the user's words as a coding task. Return only code in the language requested, no commentary, no markdown fences."
- Commit — "Convert the user's words into a Conventional Commit message. Subject under 72 chars, optional body."
- Slack — "Convert the user's words into a friendly Slack-tone message. Short. No greeting if it's a thread reply."
- Translate-DE / Translate-EN — "Translate the user's words into German / English. Return only the translation."
Activate by saying the profile name first: "Code, write a Python function that parses Apache log lines into structured records." The clean function appears at the cursor.
Accuracy & latency tips
Push-to-talk beats voice activity detection
Hotkey dictation that captures only while the key is held produces dramatically cleaner audio than always-listening tools. The transcription is more accurate because there is no leading or trailing silence to confuse the model.
Pick the right Whisper tier
Whisper Tiny / Base are fast but make more mistakes on names and technical terms. Whisper Small / Medium are the sweet spot for most laptops. Whisper Large-V3 is the highest-accuracy tier and the one to pick when correctness matters more than instant transcription.
Use streaming inference for the AI step
Most providers stream the response token-by-token. A good voice-prompting tool types the response as it streams, so the perceived latency is the time-to-first-token (often under a second), not the time-to-completion.
Don't fight pronunciation — use post-processing
For brand names, ticker symbols, and uncommon terms, configure a small replacement list rather than trying to retrain the model. "Open AI" → "OpenAI", "G P T four" → "GPT-4", "claude" → "Claude" — a handful of these covers 90% of the cases that would otherwise look wrong in the output.
Privacy model — what leaves your machine, and what doesn't
The privacy model of a voice-prompting setup splits cleanly across the two halves of the pipeline:
- Speech-to-text: with a local Whisper, audio never leaves the endpoint. There is no third party in the speech path.
- AI inference: the transcribed text reaches whichever provider you point the tool at. With cloud providers (OpenAI, Anthropic, Google), the text goes to that provider under their data policy. With a local model server (Ollama, LM Studio), nothing leaves the endpoint.
For a deeper treatment of the regulated-industry side of this, see our companion article on enterprise offline voice recognition.
A 5-minute setup
-
Install AirTypes
Linux and macOS today, Windows in development. Get the build for your OS.
-
Add your AI key
Settings → My Agent → paste an OpenAI, Anthropic, Gemini, OpenRouter, Groq, or Ollama endpoint and key.
-
Create 3 profiles
Start with Email, Reply, Summary. Add more as patterns emerge in your day.
-
Bind your hotkey
Pick a combo that doesn't conflict with your apps — Ctrl+Alt+Space works on most setups.
-
Voice-prompt
Place the cursor anywhere, hold the hotkey, say "Email, tell the team standup moves to 11", release. The composed email appears at the cursor.
FAQ
Can I voice-prompt ChatGPT, Claude, and Gemini from one tool?
Yes. Any AI provider that exposes an OpenAI-compatible chat completions endpoint works — and at this point that is essentially every major provider, either directly or via OpenRouter. With a single tool like AirTypes' My Agent, switching from GPT-4 to Claude Opus to Gemini 1.5 is a dropdown change, not a workflow change.
Is voice prompting safe for sensitive content?
The transcription step can run fully offline with Whisper, so the audio never leaves your machine. The AI inference step depends on which provider you point the tool at — cloud providers receive the transcribed text, local providers do not. For confidential content, route through a local Ollama or LM Studio model.
How does voice prompting compare to ChatGPT's built-in voice mode?
ChatGPT's voice mode is excellent for conversation inside the ChatGPT app. Voice prompting via a desktop tool wins when you want the answer to land in another app — your editor, your email client, your chat tool — without switching context. Different tools, different jobs.
Do I need to learn prompt engineering to voice-prompt well?
No. Profiles do most of the prompt-engineering work for you. You write the system prompt once for each pattern (Email, Code, Reply, Summary), then trigger it by saying the profile name and speaking naturally. The model receives a structured prompt without you having to think about it.
What's the latency like end-to-end?
Whisper transcription is sub-second on most modern hardware. Time-to-first-token from a fast provider (OpenAI, Anthropic, Groq, Gemini Flash) is usually under a second. Total perceived latency from "release the hotkey" to "answer is appearing at the cursor" is typically 1–2 seconds.
Can I use voice prompting in Microsoft Word or Outlook?
Yes. Voice prompting works in any desktop app because the AI's reply is typed at the cursor via OS keyboard injection — Word, Outlook, Excel, Teams, Slack, VS Code, your browser, your terminal. See our companion guide on Microsoft Word voice dictation alternatives.
Try voice prompting free for 7 days
BYOK to ChatGPT, Claude, Gemini, Groq, OpenRouter, or local Ollama. Audio stays on your machine.
Download AirTypes