Choosing Models
Choose transcription and enhancement models for your language, Mac, and workflow.
There are two separate choices: a transcription model turns audio into text, and an optional enhancement model processes that text. Select both in a Mode.
Transcription Starting Points
| Need | Starting point |
|---|---|
| Local dictation on Apple Silicon | Parakeet V3, the model used by onboarding. |
| English-only local dictation | Parakeet V2 or Parakeet Unified. |
| Other local languages or model behavior | Compare Parakeet Ultra, Nemotron, Whisper, Cohere Transcribe, and SenseVoice against their supported languages. |
| Native Apple transcription | Apple Speech on macOS 26 or later, with the required language assets. |
| Intel Mac | A cloud transcription provider; the app warns that local models do not work reliably on Intel. |
| Text appearing while you speak | A model supporting real-time transcription. |
| Existing audio or video files | A model supporting file/batch transcription; Cartesia Ink 2 is streaming-only. |
A model being multilingual does not mean it supports every language. Check the Mode's Language picker before downloading or connecting it. Speed and accuracy vary with your microphone, language, recording, and hardware.
See Local Models for the full on-device catalog and Cloud Providers for provider capabilities.
Enhancement Choices
| Need | Option |
|---|---|
| Local dictation cleanup | VoiceInk Refine, on Apple Silicon with at least 16 GB RAM. |
| Local custom prompts and context | Ollama with a locally hosted model. |
| Hosted cleanup, rewriting, or Assistant responses | Groq, Cerebras, Gemini, OpenAI, OpenRouter, Anthropic, or Mistral. |
| An existing command-line workflow | Local CLI. Its actual processing location depends on the command. |
| An OpenAI-compatible service | Custom Models. |
Refine has a fixed cleanup task. It does not support custom prompts, context, or Respond output. For those workflows, select a compatible enhancement provider. Dictionary Auto Learn runs locally without an enhancement provider.
Supported built-in enhancement providers offer a Custom… model option when available. Enter the exact model ID supported by your provider. OpenRouter loads its model catalogs using Refresh Models.
Compare Your Own Results
Use History → Analyze for saved recording timings or Dashboard → View Insights → Model Performance for summaries. Disable enhancement temporarily to distinguish transcription delay from enhancement delay.
Provider costs, quotas, and model availability are controlled by the provider and can change. VoiceInk's Mac license does not include third-party API usage. Consult your provider's account and billing information before choosing a hosted model.