Local Whisper voice input for your Windows AI overlay
Press Alt+V, speak, and press Alt+V again to send. AuraCloak turns your speech into a prompt with one of five engines. Choose Local Whisper and transcription runs on your PC with the base (141 MB) or small (465 MB) model, so your audio never leaves the machine.
Push-to-talk into the AI, from your own microphone
Voice input in AuraCloak is narrow on purpose. It records your microphone, only between two presses of Alt+V, and turns what you said into a chat message for the model you picked. A red pulsing indicator shows while it is listening. Alt+Shift+V throws the recording away.
It never taps speaker output or call audio. That is a deliberate privacy choice: AuraCloak can't hear the other people on your Zoom call, so there is nothing of theirs to store, transcribe or send. If you want the AI to know what someone said, you say or type it yourself.
This also makes it different from Whisper dictation tools such as LocalWhisper or Privacy Scribe, which type text into whatever window has focus. AuraCloak's voice input goes into the AuraCloak chat and gets an answer back, so it works as a spoken question box for your AI rather than a keyboard replacement.
One practical warning. Your meeting app uses the same microphone. If you are unmuted in Teams and press Alt+V, everyone hears the question. Mute first, or type.
Five speech engines, and where your audio goes
Pick one in Settings > Voice, or with Engine= in the [Voice] section of auracloak.ini.
| Engine | Where audio goes | Cost | Setup |
|---|---|---|---|
| Windows built-in (default) | Stays on your PC | Free | None. Quality depends on the Windows speech model installed. |
| Local Whisper | Stays on your PC | Free | Model files from the installer checkbox or Scripts\download-models.ps1 |
| OpenAI Whisper API | OpenAI | About $0.006 per minute | OpenAI API key |
| Google Speech-to-Text | Google Cloud | About $0.009 per minute | Google Cloud key with Speech-to-Text enabled |
| Azure Speech | Microsoft Azure | Billed per audio minute | Speech resource key and region |
Two models, both offline
AuraCloak runs Whisper through Whisper.net and NAudio. After the model file is installed, it needs no internet connection and no API key.
- base,
ggml-base.bin, 141 MB. Faster, good accuracy. Start here. - small,
ggml-small.bin, 465 MB. Slower, better with names and jargon. - The installer checks "Install offline speech recognition models (Whisper)" by default, about 600 MB for both
- Set
Engine=LocalWhisperandWhisperModel=smallinauracloak.inito switch without opening Settings
When speaking beats typing
Long questions
A 60-word question with context takes about 20 seconds to say, and much longer to type with one eye on a call.
Practicing answers out loud
Speak a full answer for interview prep or a talk rehearsal and get feedback on structure and length. Local Whisper keeps the practice audio on your PC.
Notes right after a call
Dictate what happened and ask for a ticket summary or a follow-up email. See support calls.
Less typing
If long typing sessions strain your wrists, two key presses and a spoken question replace a paragraph of typing.
What spoken prompts look like once transcribed
Transcripts and replies are examples. Accuracy depends on the engine, model and microphone.
You (Alt+V): "Turn this into action items. Priya owns the vendor shortlist by Friday, Marco checks the contract renewal date, and I send the budget sheet tomorrow."
Reply: - Priya: vendor shortlist, due Friday
- Marco: confirm contract renewal date
- You: send budget sheet, tomorrow
You (Alt+V): "Here is my answer to why I want this role..." (90 seconds)
Reply: Good opening. You list four reasons; keep the two tied to their product. The ending trails off, so finish on the team you would join.
You (Alt+V): "What's the PowerShell command to see which process is using port 8080?"
Reply: Get-Process -Id (Get-NetTCPConnection -LocalPort 8080).OwningProcess
You (Alt+V): "Can you draft a reply to... actually never mind"
Then: Alt+Shift+V. The recording is discarded and nothing is sent.
Set up local Whisper voice input on Windows
Install with models
Leave the Whisper models box checked in the installer, or run Scripts\download-models.ps1 later.
Choose the engine
Settings > Voice, pick Local Whisper and the base or small model.
Set the language
Use a language tag such as en-US, en-GB or de-DE. The default is en-US.
Talk
Alt+V to start, Alt+V to send, Alt+Shift+V to cancel.
Whisper dictation apps, ChatGPT voice and AuraCloak
| Whisper dictation apps (LocalWhisper, Privacy Scribe) | ChatGPT voice mode | AuraCloak voice input | |
|---|---|---|---|
| Where the text goes | Typed into any focused window | A ChatGPT conversation | The AuraCloak chat, sent to your chosen model |
| Offline transcription | Yes | No | Local Whisper or Windows speech |
| Push-to-talk hotkey | Yes | In-app button | Alt+V |
| Spoken replies | No | Yes | No, text replies |
| Choice of AI model | Not an AI chat | OpenAI only | OpenAI, Claude, Gemini, Copilot, Azure, local |
| Window hidden from screen share | Not applicable | No | Yes |
Ask your AI out loud, with audio that stays on your PC
$9 a month or $49 once. Voice input and local Whisper are in both plans.
Voice input questions
Does AuraCloak listen to my calls or system audio?
No. Voice input records only your microphone, and only after you press Alt+V. It never captures speaker output or call audio, so it can't hear the other people on a call.
Does Local Whisper send my audio anywhere?
No. Local Whisper runs the Whisper model on your PC through Whisper.net. Once the model file is installed it needs no internet connection and no API key. The transcribed text then goes to whichever AI provider you chose, or stays local too if you use a local model.
Which Whisper model should I choose?
Start with base (141 MB), which is the faster of the two. Switch to small (465 MB) if you see mistakes on names or technical terms; it is more accurate and somewhat slower. Set it in Settings > Voice or with WhisperModel= in auracloak.ini.
Will people on my call hear me talking to the AI?
Yes, if your microphone is unmuted in the meeting app. AuraCloak and Zoom or Teams use the same microphone. Mute yourself in the call before pressing Alt+V, or type the question instead.
Can I cancel a recording without sending it?
Yes. Press Alt+Shift+V while recording to cancel. Nothing is transcribed or sent.