Recommended Models
Our picks for fast local dictation, cloud transcription, and AI enhancement.
VoiceInk's Recommendation
Transcription
- Local
- Parakeet Ultra
- Cloud
- ElevenLabs Scribe V2
Enhancement
- Groq or Cerebras
- Qwen 3.8 27B or GPT-OSS 120B
- Anthropic
- Claude Haiku 5.5
- OpenAI
- GPT-5.6 Luna
Choose transcription and enhancement separately in a Mode: transcription turns speech into text; optional enhancement cleans up, formats, or rewrites that text.
Recommended Models for Transcription
Local transcription
Parakeet Ultra is our first pick for local dictation on Apple Silicon. It is a post-trained Parakeet V3 model covering English and supported European languages. Audio stays on your Mac after setup.
| Model | Why choose it | Real-time text |
|---|---|---|
| Parakeet Ultra | Start here for English and supported European languages. | Optional |
| SenseVoice Small | A focused option for Mandarin, Cantonese, Japanese, Korean, and English. | No |
| Cohere Transcribe | Compare it when you need another local multilingual option, including Arabic and Vietnamese. | No |
| Whisper Large v3 Turbo | A multilingual alternative when the other local picks do not give you the results you need. Processes audio after you stop. | No |
Check the Mode's Language picker before downloading: multilingual models support different sets of languages. For live text in languages outside Parakeet's coverage, also consider Nemotron Multilingual. See Local Models for the full catalog and language details.
On an Intel Mac, start with a cloud provider. VoiceInk warns that local models do not work reliably on Intel.
Cloud transcription
ElevenLabs Scribe V2 is our top cloud pick. Compare Deepgram Nova 3 for live dictation, Soniox V5 for lower usage costs, or Groq Whisper Large v3 Turbo for inexpensive batch transcription. AssemblyAI Universal-3.6 Pro is another real-time option.
Select Scribe V2 in your Mode and enable Real-time for live text; VoiceInk uses the corresponding streaming model automatically. Scribe V2 Medical is a separate, batch-only choice for clinical terminology.
Cloud cost comparison
These estimates assume two hours of transcription per day. Batch processes a completed recording; real-time returns text while you speak.
| Provider and model | Batch / day | Real-time / day |
|---|---|---|
| Groq · Whisper Large v3 Turbo | $0.08 | Not supported |
| Soniox · V5 | ≈$0.20 | ≈$0.24 |
| Deepgram · Nova 3 | See provider rates | $0.92–$1.10 |
| ElevenLabs · Scribe V2 | $0.44 | $0.78 |
| AssemblyAI · Universal-3.6 Pro | Uses 3.5 Pro or Universal-2 | $0.90 |
| AssemblyAI · Universal-3.5 Pro | $0.42 | See provider rates and model availability |
| AssemblyAI · Universal-2 | $0.30 | Not supported |
Prices checked November 3, 2026, in USD before credits, taxes, and add-ons. Deepgram's range covers monolingual and multilingual pay-as-you-go streaming. Soniox uses approximate hourly equivalents for token billing. ElevenLabs shows advertised hourly rates; check plan requirements and included usage.
For files or batch recording, VoiceInk routes Universal-3.6 Pro to Universal-3.5 Pro or Universal-2, depending on the selected language. Some languages are available only in real-time. See Cloud Providers for the full catalog. Your VoiceInk license does not include provider usage.
Recommended Models for Enhancement
Start with Qwen 3.8 27B or GPT-OSS 120B on Groq or Cerebras. For Anthropic or OpenAI, choose Claude Haiku 5.5 or GPT-5.6 Luna. All listed hosted models are available in VoiceInk's enhancement picker.
| Provider | Recommended model | Why choose it |
|---|---|---|
| Cerebras | qwen-3.8-27b or gpt-oss-120b | Fast cleanup and custom prompts. Qwen is the app's default for this provider. |
| Groq | qwen/qwen3.8-27b or openai/gpt-oss-120b | The same model families on Groq, with different provider-specific IDs. Qwen is the default. |
| Anthropic | claude-haiku-5-5 | Claude Haiku 5.5 for everyday enhancement. Select it explicitly; the app defaults to Sonnet 5.5. |
| OpenAI | gpt-5.6-luna | GPT-5.6 Luna for everyday enhancement. Select it explicitly; the app defaults to GPT-6 Luna. |
| Gemini | gemini-3.5-flash-lite | A starting point for lightweight cleanup. Try gemini-3.8-flash for more involved rewriting or prompts. |
| VoiceInk Refine | VoiceInk Refine V1 | Our local pick for fixed dictation cleanup on Apple Silicon with at least 16 GB RAM. No API key required. |
For other hosted models, choose Custom… and enter an ID supported by your provider. With OpenRouter, use Refresh Models to load its enhancement catalog.
For private cleanup, use VoiceInk Refine. It supports Paste and Custom Command output, but not custom prompts, context, Respond output, or Auto Learn review. For local custom prompts, connect Ollama to a locally hosted model.
Aim for enhancement under two seconds for short, everyday dictation. If it regularly takes longer, compare another model or provider. Longer recordings and complex prompts can take more time.
Set Up Your Picks
- Download a local model under AI Models → Local, or add and verify your API key under AI Models → Cloud.
- Edit a Mode, choose its transcription model and language, and enable Real-time if you want streaming on a supported model.
- For enhancement, enable AI Enhancement, then select the provider and exact model from the table above. Leave enhancement off if you only need the transcript.
For imported audio or video, choose a batch-capable model. Cartesia Ink 2 is streaming-only. See Real-time Transcription and Transcribe Audio Files.
Compare Your Own Results
Use History → Analyze for recording timings or Dashboard → View Insights → AI Model Performance for summaries. Temporarily disable enhancement to separate transcription delay from enhancement delay.
Compare a few recordings with your usual vocabulary and speaking style. Keep the setup that needs the least editing with an acceptable wait.