When dictation doesn't work
Dictation has two ways of turning speech into text Since 1.20. Normally the text appears while you speak: that is OpenAI’s live transcription. When it can’t be reached, or breaks off, Snotra keeps recording and, when you stop, has the rest transcribed in one piece by OpenAI’s Whisper — the way dictation worked before. Either way the text ends up in the input field; the second way you only see it at the end.
Both need an OpenAI key under Settings › Models. Without one there is no dictation at all.
What you need
Section titled “What you need”A recording that did not go as you expected. The line next to the microphone says what happened — read it before anything else.
How to tell what is happening
Section titled “How to tell what is happening”| What you see | What it means |
|---|---|
| The text appears while you speak, grey and slanted. Next to the microphone: Recording …, a red level and a clock. | Live text works. |
| Recording without live text — the text follows when you stop. | The live transcription failed. Snotra goes on recording; what was already written stays. Keep talking. |
| After the stop, Transcribing with Whisper …, then No live text (…). Transcribed with Whisper instead. | The fallback did its job: the rest of the recording is in the field. The reason in brackets says why live text failed. |
| After the stop, Error: … | Whisper failed as well. What was in the field stays; the part of the recording that had no text yet is lost. |
| Error: No OpenAI key stored (voice input needs one). right after the click | There is no key, and the microphone did not even switch on. |
| The level next to the microphone stays flat while you speak. | The microphone hears nothing — see If it doesn’t work. |
When the live text fails
Section titled “When the live text fails”The reason is in the brackets of No live text (…).
| The reason | What it means | What to do |
|---|---|---|
| Could not reach OpenAI’s speech recognition. Check the OpenAI key and the network. | The live connection did not come about, or broke off. Most often a firewall, proxy or VPN lets ordinary requests to OpenAI through, but not the live connection (a WebSocket to api.openai.com). A key OpenAI refuses looks the same from here. | Nothing urgent — the fallback covers it. To get live text back: check the key under Settings › Models, try without the VPN, or ask whoever runs the network to allow WebSocket connections to api.openai.com. |
| An English sentence from OpenAI, such as … does not exist or you do not have access to it | Your OpenAI account may not use the live model, gpt-live-transcribe, yet. With OpenAI some models need a verified organisation. | Check the organisation and its limits in your OpenAI account. Until then the fallback transcribes every recording. |
| Rate limit reached, insufficient_quota | As with the chat: too many requests, or no credit left. | See When the model reports an error. |
When Whisper fails too
Section titled “When Whisper fails too”Error: … after the stop means neither way got through. The text in the field stays; say the missing part again once the cause is gone.
- Connection to the provider failed., fetch failed, Timed out after … s. No network to OpenAI at all. Check the connection.
- Incorrect API key provided, insufficient_quota. The key or the account: see When the model reports an error.
- No speech recognised. The recording was silent or too quiet — see below.
What happens
Section titled “What happens”Nothing you dictated is thrown away by a failure: whatever is in the input field stays there, as ordinary text, and you can change or send it. A failed live connection costs nothing extra; the fallback sends only the part of the recording that has no text yet.
If it doesn’t work
Section titled “If it doesn’t work”- Microphone access denied. Your system does not let Snotra use the microphone. On macOS switch on Snotra Agent under System Settings › Privacy & Security › Microphone; on Windows allow desktop apps under Settings › Privacy & security › Microphone. Then try again.
- The level stays flat, and no text comes. The microphone hears nothing. Snotra records from the system’s default input: check in the sound settings that it is the microphone you speak into, and that it is not muted.
- Sentences break in odd places. A sentence ends at a pause of about 0.7 seconds. A long breath in the middle of one ends it there; nothing is lost, only the full stop may sit in the wrong place. Correct it in the field.
- The recording stopped by itself. A recording lasts five minutes at most; half a minute before, the line next to the microphone says so. Start a new one.
- The recording ended when the window was minimised. Hiding the window ends a recording, and the text so far stays. Keep the window open while you dictate.
Cmd+D/Ctrl+Ddoes nothing. The shortcut works while the Snotra window is in front.