In short
When on-device speech recognition produces nothing on a Mac, check four things in this order: the macOS version and the chip, the language the recogniser is set to, whether the speech model for that language has finished downloading, and whether there is any audio to recognise in the first place.
This guide covers each, with the case of Notey, which transcribes meetings with Apple's on-device recogniser, at the end.
1. The Mac and macOS version
Apple has more than one speech recognition system. The one apps have used for years, SFSpeechRecognizer, runs on many Macs. The newer one, SpeechAnalyzer, arrived in macOS 26 and is designed for long audio such as meetings and lectures. An app built on the newer framework needs macOS 26 or later.
To check your version, choose Apple menu › About This Mac. The same window shows the chip: an Apple M-series chip is Apple silicon; anything that names Intel is not.
If you are on an Intel Mac, see can you transcribe meetings on the device on an Intel Mac for what is and is not possible.
2. The language
A speech recogniser listens for one language at a time. If it is set to English and the meeting is in German, the transcript will be empty or nonsense, and nothing will look broken.
Check two things:
- Which language the app is using. Some apps have a setting. Others use the language your Mac is set to, in System Settings › General › Language & Region.
- Whether that language is supported. On-device recognition covers a limited set of languages. Which languages can a Mac transcribe on the device lists them.
A meeting that switches between languages is transcribed in the one the recogniser was set to. The parts in the other language will be poor.
3. The speech model
On-device recognition needs a speech model for the language, stored on the Mac. macOS downloads it on request, once. Until it has finished, the recogniser cannot run.
What you see depends on the app. Apps that use the newer framework can ask macOS to download the model and report progress; others simply fail until it is there. If transcription does not start on a new Mac, or after changing your Mac's language, a model that has not finished downloading is the most likely cause.
To help it along:
- Stay connected to the internet until the app reports the model is ready.
- Make sure the Mac has free disk space.
- If the app offers a download button, use it rather than waiting for a recording to trigger it.
Once the model is on the Mac, recognition runs without a network connection.
4. The audio
A recogniser that hears silence writes nothing. Before blaming speech recognition, check the recording has sound: a flat waveform or a silent playback means the problem is the microphone, a permission, or the input device. Meeting recording silent or transcript empty covers those.
A recording that is audible but very quiet, distant or full of cross-talk gives a thin, patchy transcript rather than an empty one. That is a quality problem, not a failure.
Dictation is a separate check
macOS Dictation, switched on in System Settings › Keyboard, is a useful quick test of the microphone. Apple's documentation notes that you can check in Keyboard settings whether Dictation's voice input is processed on your device. (Apple: dictate messages and documents on Mac)
Dictation working proves the microphone works and the Mac can recognise speech. It does not prove a particular app's recogniser has its language model, or is set to the right language.
In Notey
Notey transcribes with Apple's on-device speech recognition, live, while it records. It needs macOS 26 or later on Apple silicon and will not open on anything earlier.
Notey transcribes in the language your Mac is set to. If your meetings are in a different language from the one in System Settings › General › Language & Region, the transcript will be poor. Live transcription covers ten languages: English, French, German, Italian, Spanish, Portuguese, Japanese, Korean, Chinese and Cantonese.
Notey checks at launch rather than at the end of your first meeting, and shows a notice if something is missing:
- "The speech model has not been downloaded yet", with a Download button. "Notey needs it once, and then transcribes without a network connection. Recording works in the meantime." Press Download and stay online until the notice disappears.
- "Your language is not supported for on-device transcription." Notey will not send your audio anywhere else to transcribe it. Recordings are still made and kept.
In every case, recording carries on and the audio stays on your disk, so no meeting is lost while transcription is unavailable.
Two separate, stronger recognisers re-transcribe English meetings after they end. They are about 1.2 GB in total, fetched once in the background, and never downloaded during a recording. Until they have arrived, you keep the live transcript, and the meeting is improved later. Why Notey re-transcribes a meeting after it ends explains it.
Frequently asked questions
Does on-device speech recognition work offline on a Mac?
Yes, once the speech model for your language is on the Mac. The model is downloaded once, which needs a network connection; after that, transcription runs without one.
Why is my transcript in the wrong language or full of nonsense?
The recogniser is listening for one language and the meeting was in another. Apps that use Apple's recogniser choose a language, often the one your Mac is set to, and speech in a different language comes out garbled or not at all.
Which Macs can use Apple's newest on-device speech recognition?
Apple's SpeechAnalyzer framework arrived in macOS 26. Notey, which uses it, requires macOS 26 or later on a Mac with Apple silicon.
Why is dictation working when an app's transcription is not?
Dictation and an app's transcription can use different speech models and different language settings. Dictation working shows the microphone is fine; it does not show the app's language model is installed.