VOICE LAB
Find a voice that fits.
One Agent OS interface for speech services across AI providers. Provider access and spending policies belong behind that interface.
Integration pending: this page does not yet connect Grok or Codex. The intended test is to send and receive speech through real provider adapters here. Current device diagnostics and manual observations are supporting tools, not completion of that goal. Claude is excluded from current testing; no paid API calls are connected.
Your reference sample
Read this aloud for dictation. Use it for read-aloud comparisons too. Repeat each candidate three times in the same room and on the same device.
Use a harmless sample in native agents: submitting a dictated instruction can start work and consume your normal allowance.
Local read-aloud
Only voices the browser identifies as local appear here.
Checking this browser…
Measures time to the browser’s speech-start event, not acoustic latency. No automatic remote-voice fallback.
Microphone & dictation
First allow access, then choose a microphone and speak to check its input level.
Access has not been requested. Click Enable microphone and allow access in your browser.
The input check runs locally for up to 30 seconds. Audio is not recorded or uploaded. If the app browser blocks access, open this URL in your regular browser and check its microphone permissions.
On-device dictation
Experimental browser recognition. Requires an installed local English language pack and microphone permission.
Dictation support has not been checked. Microphone access and local recognition support are separate.
Dictation uses the browser/OS default microphone; the selector above applies to the input check only. Set your preferred default in browser or system settings before dictating. Stops after 30 seconds. If a language pack is missing, this first experiment reports that requirement; it does not download one.
Native-client observations are reference only.
These instructions can supply diagnostic comparisons. They are not the intended Agent OS workflow or proof of a working unified adapter. Any manual test must stay within your authorized allowance and spending policy.
| Candidate | Where and how to test manually | Agent OS access |
|---|---|---|
| OpenAI / ChatGPT / Codex | In the ChatGPT desktop app, open a Codex task and select Start voice chat where available. Allow its microphone access. Availability depends on account, rollout and workspace. Official OpenAI instructions. | Experimental Codex realtime protocol found. Account access, support and billing still unverified. |
| Grok | Open Grok or its mobile app and use its voice feature where available. Try the same short exchange within your included allowance; note shared quota before and after. Official Grok guide. | Consumer voice is separate from an embeddable, qualified subscription adapter. |
| Cursor | Dictate into the native composer; inspect without sending. | Native dictation documented. Reusable interface and token accounting unverified. |
| Antigravity | Use interactive CLI /voice or /record; inspect before sending. | Interactive-only path; headless interface and token accounting unverified. |
| whisper.cpp / Piper | Local transcription / local speech synthesis candidates. | Not installed or connected in this experiment. |
Repeatable native test sequence
- Record app version, account plan, device, microphone and language. Check the allowance and extra-spend settings.
- Dictation: read the sample; paste the unedited transcript into the scorecard. Do not submit it to a working agent.
- Read-aloud: request the same words, then note pronunciation, comfort and time until the first sound.
- Live conversation, only if included: ask “What is two plus two?”, interrupt with “Stop”, then ask “What did I just ask?” Record whether interruption and context worked.
- Check usage again. Repeat three times; record failures too. Unknown or delayed usage reporting is not evidence of zero cost.
Record a test you ran in another app
These entries are your observations. Automatic device measurements are labeled separately. Results stay in this browser; export them to bring findings back to planning.
Recorded trials 0
Sample text and transcripts are saved with results. Raw microphone audio is never recorded by this page. Clearing removes this lab’s saved trials from this browser only.