recordist

Transcription models

Our recommendation: keep the default. On any Apple-silicon Mac with 8 GB of memory or more, the model Recordist offers on first launch (large-v3-turbo) gives the best accuracy, handles more than 90 languages, and keeps up with a live call. You only need this page if your machine is older or smaller, or you are curious.

Transcription always runs on your computer. The model is a file you download once, and your recordings and transcripts never go anywhere.

What the choice means for you

  • Accuracy. Larger models get names, numbers and accents right more often. The default is the most accurate option.
  • Speed. Smaller models finish sooner. The default transcribes a one-hour meeting in a few minutes on Apple silicon, and the live transcript keeps up while you talk.
  • Disk and memory. The default download is about 574 MB and uses about 1.5 GB of memory while transcribing. The smallest model is 75 MB.

If your machine is smaller

  • Less than 8 GB of memory: choose small. It is noticeably lighter, and still good for English calls with clear audio.
  • You only need to search, not read: base or tiny are fast enough for anything and fine for keyword search.
  • Meetings in more than one language: stay with the default. The smaller models are weaker outside English.

Change the model at any time under Settings → Transcription. Changing it does not re-transcribe old meetings; use Re-transcribe on a meeting if you want to.

Details

The models

Sizes are for the quantised (q5_0) or default GGML files Recordist downloads. Memory is approximate peak use during transcription. Speed is relative to real time on an Apple M2; a “10×” model transcribes a one-hour meeting in about six minutes.

Model Download Memory Speed (M2) Accuracy Use it when
tiny 75 MB ~0.4 GB ~60× Basic Very old hardware; quick drafts; you only need search
base 142 MB ~0.5 GB ~40× Fair 4 GB machines; English-only calls with clear audio
small 466 MB ~1 GB ~20× Good 8 GB machines where the live transcript matters more than perfection
medium 1.5 GB ~2.6 GB ~6× Very good Heavy accents, noisy audio, many languages; you can wait
large-v3-turbo (default) 574 MB (q5_0) · 1.6 GB (f16) ~1.5 GB ~15× Best Apple silicon and modern PCs

large-v3-turbo is a distilled version of large-v3 with most of its accuracy at a fraction of the cost.

Platform notes

  • Apple silicon (M1 or later): keep the default. Metal acceleration makes it comfortable even on 8 GB.
  • Windows with an NVIDIA GPU (early access): keep the default; CUDA is used automatically.
  • Windows or Linux on CPU only (early access): small is the sweet spot; base on laptops with fewer than four cores.
Platform Backend Notes
macOS (Apple silicon) Metal GPU acceleration, on by default
Windows CUDA → Vulkan → CPU First available backend wins
Linux CUDA → Vulkan → CPU Vulkan covers AMD and Intel GPUs

Override the backend under Settings → Transcription → Advanced if a driver misbehaves.

Live transcript and final transcript

During a meeting Recordist transcribes short chunks so you can read along. When the recording stops it re-transcribes the whole file with longer context, which fixes words the live pass got wrong. Both passes use the same model. If your machine cannot keep up during the meeting, the live transcript falls behind but the final transcript is unaffected; turn off Live in Settings → Transcription to save battery.

Language

Language is detected per meeting from the first 30 seconds of speech. Pin a language in Settings → Transcription → Language if detection is unreliable, for example when a meeting starts with a lot of names.

Managing model files

Settings → Transcription lists installed models with their size and a Remove button. Files live in the models folder inside your data folder and follow the whisper.cpp naming, for example ggml-large-v3-turbo-q5_0.bin. You can place a compatible GGML file there and Recordist will offer it.

Model downloads come from the whisper.cpp release mirror over HTTPS and are checked against a published SHA-256 before use. The app performs two other downloads, the built-in notes engine runtime and the notes model, and checks for its own updates at launch and every six hours. Every one of these appears in the Privacy Ledger.

Whisper is a model released by OpenAI under the MIT licence; whisper.cpp is an open-source project under the MIT licence. “Powered by” statements are factual and imply no endorsement.