# Command an army of agents with your voice

> Dictate prompts to Claude Code and Codex with on-device NVIDIA Parakeet or OpenAI realtime. Cost, speed, privacy, and when to drop Superwhisper.

Published 2026-09-27 on the Starkomand blog: https://starkomand.com/blog/superwhisper-alternative-for-coding-agents/

You hold a key, say "run the billing tests and tell me which ones fail", and a dictation app types it wherever your cursor happens to be. If the cursor is in a Claude Code or Codex terminal, that works. If you pay for it every month, it's worth asking what the subscription is buying you.

This post compares the two speech engines built into [Starkomand](https://starkomand.com/): NVIDIA Parakeet running on your Mac, and OpenAI's realtime transcription. It makes the case for a Superwhisper alternative for coding agents. If most of what you dictate is prompts to agents, you may not need a separate dictation subscription. I'll also say plainly when you should keep yours.

## What a dictation app leaves to you

A general dictation app turns speech into text at the cursor. That is its whole job, and the good ones do it well. With one agent open, it's enough.

With several agents, the text is the easy part. You still find the right terminal, click into it, speak, check the text and press Return. Then you do it again for the next agent. The dictation app doesn't know which window is Claude Code working on the API and which is Codex fixing the navigation, so choosing the recipient stays your job.

In Starkomand, the recipient is part of the sentence. Every agent window has a human first name, and you address it the way you'd address a person:

> Marcus, run the unit tests for the billing module and tell me which ones fail.

Hold **Left Option**, say that and let go. The words fly into Marcus's window and are submitted as his next prompt, with no click and no Return. One breath can reach two agents: "Marcus, run the tests. Diana, update the changelog" sends each half to its own window. A sentence without a name goes to the agent you spoke to last. If you say a name the app can't pin to exactly one window, it warns you instead of guessing.

The same key runs the canvas: "open two Codex agents", "close Marcus", "tidy up", "next workspace".

![The Starkomand canvas while it listens. Four agent windows, Marcus and Diana running Claude Code and Felix and Oscar running Codex; the spoken questions "Marcus how are you doing" and "Diana what's up" fly into those two windows, and a third question to Oscar leaves the voice button.](https://starkomand.com/assets/landing/voice-command-poster.jpg)

*A frame from a screen recording of the app. The questions are illustrative.*

## Two speech engines, one setting

Pick one under **Settings → Voice & speech → Voice provider**. You can switch at any time; the change applies to the next thing you say.

### NVIDIA Parakeet, on your Mac

**Parakeet · On device** runs NVIDIA's Parakeet TDT 0.6B v3, a 600-million-parameter speech model released under the CC-BY-4.0 license, through Apple's MLX framework on the Mac's own chip. It needs one download: 2.7 GB the first time, about 2 GB on disk afterwards. The download is pinned to one exact model version and checked against its digest before it is used. After that, dictation works offline.

Your audio doesn't leave the Mac. Each recording is written to a temporary file, transcribed and deleted. The model stays loaded, so the first key press after launch records straight away instead of waiting for a model to start. Keeping it loaded costs about 1.2 GB of memory.

Parakeet v3 transcribes 25 European languages, detects which one you're speaking, and adds punctuation and capital letters itself. NVIDIA reports a 6.34% average word error rate on the Hugging Face Open ASR Leaderboard.

I timed it on my own Mac, an M1 Pro with 16 GB of memory, while other agents were busy in the background. I recorded three prompts with the macOS `say` voice and transcribed each one five times:

| Prompt length | Median time from release to text |
| --- | --- |
| 4 seconds | 0.5 s |
| 16 seconds | 1.1 s |
| 34 seconds | 3.4 s (fastest run 2.0 s) |

All three came back word for word, apart from punctuation and "two hundred" written as "200". A synthetic voice is easier to recognize than yours in a real room, so read this as a speed test, not an accuracy score.

### OpenAI realtime transcription

**OpenAI · Realtime** streams your audio to OpenAI's `gpt-live-transcribe` model while you speak, so words start appearing before you let go of the key. Choose it, paste an OpenAI API key, and select **Save and use OpenAI**. The key is saved in a private file on your Mac. If you haven't saved one, Starkomand uses an `OPENAI_API_KEY` from its server's environment.

Starkomand asks OpenAI for its low-latency setting and sends a short description of the context, plus keyword hints for names such as Claude, Codex, Cursor and GLM, so product names come back spelled correctly.

Usage is billed to your own OpenAI account. At the time of writing, OpenAI lists `gpt-live-transcribe` at $0.017 per minute of audio. Audio goes out only while you hold the key, one recording is capped at two minutes, and each recording gets a fresh session, so nothing you said earlier is carried into the next one.

### Parakeet vs OpenAI realtime, side by side

| | Parakeet · On device | OpenAI · Realtime |
| --- | --- | --- |
| Where your audio goes | Stays on the Mac | Streamed to OpenAI |
| Cost per minute | None | $0.017, billed by OpenAI |
| Setup | One 2.7 GB download | An OpenAI API key |
| Works offline | Yes | No |
| Text while you speak | No, it arrives after you release | Yes |
| Languages | 25 European languages | Many; check OpenAI's list for yours |

Parakeet is the default, and the one I'd start with: it costs nothing per prompt, needs no key and keeps working on a plane. OpenAI is the better pick when you want to watch the words land as you talk, or when you dictate in a language Parakeet doesn't cover.

## What it costs next to a dictation subscription

Published prices on September 27, 2026:

- **Superwhisper Pro:** $8.49 a month, $84.99 a year, or $249.99 once for a lifetime license. Its free plan includes local Whisper models and works in any app.
- **Wispr Flow Pro:** $15 a month, or $12 a month billed yearly.
- **Parakeet in Starkomand:** no per-minute cost. It runs on the Mac you already have.
- **OpenAI in Starkomand:** $0.017 per minute of speech, billed by OpenAI.

To make the OpenAI number concrete: Superwhisper Pro's monthly price buys about 500 minutes of `gpt-live-transcribe`. Sixty 20-second prompts in a day is 20 minutes of audio. Keep that up for 22 working days and you've spoken 440 minutes, about $7.50. Most days will be well under that, and days on Parakeet cost nothing.

Starkomand itself isn't free, and I won't pretend otherwise. After the 3-day trial, Starkomand Pro is a one-time purchase for one Mac that includes a year of updates, and voice is part of it. The current price is in the [Starkomand Pro pricing section](https://starkomand.com/#pricing). The line that matters here is the recurring one: a subscription you renew every month for as long as you dictate, next to a speech engine that runs on hardware you already own.

## When to keep your dictation subscription

Starkomand is not a system-wide dictation app, and I don't want anyone to cancel something they still need.

- **It only listens inside Starkomand.** Push to talk works while the Starkomand window has focus. It won't type into Slack, email, a document or your editor. If you dictate into those every day, keep a system-wide tool.
- **Your words arrive as spoken.** There are no modes that polish a ramble into an email or reformat a paragraph. For agent prompts I prefer it that way; for prose you may not.
- **No meetings or files.** It doesn't record meetings or transcribe audio files. It transcribes the prompt you're speaking, and that's all.
- **Languages.** Parakeet covers 25 European languages. For anything else, OpenAI is your option in Starkomand, or keep the tool that already handles your language.

If you want both, one reasonable setup is Superwhisper's free plan for the rest of your Mac and Starkomand for your agents. Whether that covers your day is your call.

## Getting the most out of voice prompts

- **Say the name first, then the task.** Crew names are ordinary first names on purpose. Speech models spell common names correctly, where invented call signs come back mangled. That matters most with Parakeet, which takes no keyword hints. Starkomand also knows the usual mishearings, so "Markus" still reaches Marcus.
- **Give each agent one task, and brief several in one breath.** Each name starts a new prompt, and everything up to the next name belongs to it.
- **Speak the long briefs.** They are where talking is most comfortable. The 34-second planning brief in the timing test came back as one clean paragraph, punctuation in place.
- **Choose your own key.** Left Option is the default because a bare modifier types nothing. Pick another under **Settings → Keyboard**.
- **Go hands-free.** With hand control on, holding the "call me" sign works like holding the key.
- **Let the agents answer out loud.** With **Announce finishes and questions** on, under **Settings → Voice & speech → Spoken updates**, each agent's name is spoken when it finishes or has a question. For what to do when you come back to several finished agents, read [how to keep track of multiple AI coding agent sessions](https://starkomand.com/blog/resume-two-projects-with-agent-recaps/).

## The Starkomand features in this post

- [Voice commands that reach coding agents by name](https://starkomand.com/#scene-voice-title)
- [On-device dictation and what stays on your Mac](https://starkomand.com/#local)
- [Hand control and the rest of the canvas](https://starkomand.com/#features)
- [One-time Starkomand Pro pricing](https://starkomand.com/#pricing)
- [Questions about providers and privacy](https://starkomand.com/#faq)

If you mostly pay for dictation to talk to Claude Code and Codex, try this before your next renewal. [Start the free 3-day trial](https://starkomand.com/), download Parakeet, and speak your next few prompts by name.