Transcription
After each sync, every recording with audio and no transcript is transcribed, oldest first, and the result is stored as transcript.md in the recording’s folder:
**Speaker 1** [00:00:29]
What the first speaker said.
**Speaker 2** [00:00:30]
The reply.
Speakers are numbered in order of first appearance, per recording, until you name them.
Choosing a provider
Settings, Transcription, Speech-to-text model: a provider, then one of its models.
| Provider | Needs | Speakers | Word times | Notes |
|---|---|---|---|---|
| ElevenLabs (Scribe) | An API key | Yes, and a library of voices it recognises | Yes | The most complete: key terms, the balance on the Overview, the remote transcript deleted after saving |
| OpenAI | An API key | With gpt-4o-transcribe-diarize; whisper-1 gives one speaker |
With whisper-1; the diarising models give a time per stretch of speech |
Files up to 25 MB |
| OpenRouter | An API key | Depends on the model | Required: a model that gives no times fails the recording | Files up to 25 MB; key terms where the model takes them |
| Local (Whisper on this Mac) | A 1.7 GB download | Yes, per recording | Yes | No key, nothing leaves the Mac |
Changing the provider, the model or the language transcribes nothing again. To transcribe recordings again, see below.
ElevenLabs
In the ElevenLabs web app, create an API key restricted to the speech-to-text and user-read permissions (speech_to_text and user_read). Paste it into the ElevenLabs row of API keys in Settings and press its Save; it is checked with ElevenLabs before it is stored.
The models are scribe_v3, scribe_v2 and scribe_v2_medical. Its settings:
- Key terms, one per line: names and words the model should know. Listing any adds 20% to the cost.
- Warn below so many hours of transcription left.
- Delete the transcript at ElevenLabs after saving, on by default.
- Recognise speakers from the ElevenLabs speaker library; see Speakers.
Cost
Transcription is billed by audio length, from the ordinary credit balance. On the Starter plan in November 2026 it cost 1,000 credits per hour of audio. The Overview shows an estimate of the hours left; it is an estimate, and ElevenLabs’ refusal of an upload is what decides that the balance is used up.
A pass stops starting new uploads after 20 minutes; the rest follow at the next sync.
OpenAI and OpenRouter
Store the provider’s key in API keys and choose it as the speech-to-text provider. The answer is stored in the same format as Scribe’s, so the transcript, the speakers, cleanup and summaries work the same way.
- OpenAI offers
whisper-1, which gives the time of every word but no speakers, and the models whose ID containstranscribe-diarize, which tell the speakers apart per stretch of speech. OpenAI’s other transcription models give no times at all and are not offered. - OpenRouter lists its transcription models. The app asks for word times, speakers and the key terms; a model that cannot tell speakers apart or take key terms is asked once more without them.
- Both refuse a file over 25 MB, about two hours of Plaud’s audio. Such a recording fails before anything is sent; the app does not split or re-encode audio.
- There is no balance card, low-balance warning or speaker library for them.
On this Mac
With Local as the provider, recordings are transcribed on the Mac by Whisper through WhisperKit, and the speakers are told apart by SpeakerKit, both in Core ML. No key is needed and nothing of the recording leaves the Mac.
Press Download next to the model’s files in Settings: about 1.7 GB from Hugging Face, with the progress shown. Cancel keeps the files already complete, and Download again fetches only what is missing. Remove deletes the downloaded models. The app never downloads anything by itself.
- Until the files are complete, recordings wait: no attempt is counted and nothing is marked failed.
- The first transcription after a download is slow, because Core ML prepares the model for this Mac, which takes a few minutes.
- After that, an hour of audio takes about 4 minutes on an M4. The app uses about 1 to 1.5 GB of memory while it transcribes, plus about 230 MB per hour of the recording.
What you give up against Scribe: accuracy (Scribe was found clearly better on Czech), the speaker library (speakers are named per recording) and key terms.
The model is large-v3-v20240930_turbo. Its files and their licences: the Whisper model in Core ML, argmaxinc/whisperkit-coreml (MIT); its tokenizer, from openai/whisper-large-v3 (Apache-2.0); the speaker models, argmaxinc/speakerkit-coreml (CC-BY-4.0).
Transcribing again
In Recordings, select the recordings and press Transcribe again…. The footer and the confirmation show the number of recordings, the hours of audio, and about how many credits it costs or, on this Mac, about how long it takes. A hosted provider charges it like a first transcription.
The old transcript is overwritten, and with it the cleaned text, your corrections and the names you gave “Speaker N”. The audio stays. The new transcript is cleaned and summarised again when models are chosen.
Cancel stops after the current recording. A recording not done in this run shows “Queued” and is done by the next syncs; to drop it instead, select it and press Remove from queue.
Turning it off
Turn off Transcribe recordings in Settings, or remove the key. The audio is still synced.