GuidesMModels & BYOKChoosing a speech model
Beginner6 min read

Choosing a speech model

On-device versus cloud, which Whisper build to download, and what GPU acceleration is actually worth.

Transcription can run on your machine or in the cloud. The trade-off is the obvious one — privacy and offline capability against accuracy and speed on modest hardware.

SettingsAI Models

On-device models

These are Whisper builds, downloaded on demand and stored locally. Six are offered:

ModelSizeNotes
Small466 MBThe only model available on the free plan; usable on CPU
Large V3 Turbo1.62 GBRecommended — near-Large accuracy, much faster
Large V3 Turbo Q5574 MBQuantised, smallest of the Turbo family
Large V3 Turbo Q8874 MBQuantised, balanced
Medium1.53 GBSuperseded by Turbo for most uses
Large V33.09 GBMost accurate, slowest
!
Medium and larger really want a GPU. On CPU they transcribe slower than you speak, which makes dictation useless rather than merely sluggish. On a CPU-only machine, stay on Small or use a cloud model.
Everything except Small requires Pro or a perpetual licence. A separate Qwen3-ASR engine is also offered on Pro, with wider language and dialect coverage, in 0.6B (~879 MB) and 1.7B (~2.4 GB) builds.

GPU acceleration

GPU use is on by default where it is supported. What that means depends on your platform:

PlatformAccelerationSetup
macOS (Apple Silicon)Metal, roughly 7× faster than CPUNothing to do
Windows (NVIDIA)CUDA, dramatically fasterA GPU runtime is downloaded on request from the On-Device Models page
Windows (other) · LinuxCPUThe toggle is disabled when no capable device is found
The Windows CUDA runtime is not bundled — it is a separate download of roughly 160 MB, plus around 580 MB of NVIDIA libraries if you do not already have the CUDA toolkit installed. You need a working NVIDIA driver, and a restart of the app afterwards.

On macOS you can additionally download a Neural Engine encoder for a model from the same page. It is never installed automatically, and if it fails to load, transcription silently continues on Metal.

Cloud models

Cloud transcription is more accurate on accented speech, noisy rooms and technical vocabulary, and it does not care how fast your machine is. Signing in gives you a built-in provider with no key setup — see Using the built-in provider.

Mixing the two

You do not have to choose once. Each workspace can override the speech model, so a workspace for confidential notes can pin an on-device model while everything else uses the cloud. Set it on the workspace, not globally.

What happens when the cloud fails

On a network error or timeout, InkSpoke falls back to a local model and finishes the job. Two exceptions are worth knowing:

  • Quota, billing and authentication errors do not fall back — retrying locally would hide a problem you need to fix.
  • On the free plan there is no local fallback, because on-device models beyond Small are not part of that plan. The recording is preserved instead and offered with a Retry action, so your words are not lost.
Next in this track

Keep going.