Custom Models
Neuron Flame supports two local model families and any OpenAI-compatible cloud transcription endpoint. Pick what fits your hardware, language, and accuracy needs.
Parakeet V2 (default)
- Made by: NVIDIA, ported to Apple Silicon by FluidAudio.
- Size: 474 MB
- Languages: English only
- Speed: Fastest. Runs on the Apple Neural Engine.
- Accuracy: Excellent for English, comparable to Whisper Medium.
- Best for: English speakers on Apple Silicon Macs.
Whisper family
OpenAI's Whisper, ported to whisper.cpp by Georgi Gerganov. Multiple sizes:
| Model | Size | Speed | Accuracy | Languages |
|---|---|---|---|---|
| Tiny | 78 MB | Fastest | Lowest | 100+ |
| Base | 148 MB | Very fast | Good | 100+ |
| Small | 244 MB | Fast | Better | 100+ |
| Medium | 770 MB | Moderate | Great | 100+ |
| Large v3 | 1.6 GB | Slower | Best | 100+ |
| Large v3 Turbo | 815 MB | Fast | Excellent | 100+ |
The Turbo variant is a special case: it's a distilled Large with comparable accuracy at half the size and twice the speed. Often the right choice if Parakeet's English-only is a dealbreaker.
Downloading a Whisper model
- Open the AI Models tab.
- Find the model in the list.
- Click Download. Progress shows in the row.
- Once complete, click Use this model.
Models live in ~/Library/Application Support/com.neuronflame.app/Models/. You can delete unused ones to reclaim space.
Custom OpenAI-compatible endpoints
Want to use a hosted Whisper API (e.g. Groq's blazing-fast Whisper Large), a self-hosted Whisper server, or any other endpoint that speaks OpenAI's /v1/audio/transcriptions protocol? Add it as a Custom Cloud Model.
- AI Models tab → + Add Custom Provider.
- Enter a name (e.g. "Groq Whisper").
- Base URL — e.g.
https://api.groq.com/openai/v1. - API key.
- Model name — e.g.
whisper-large-v3-turbo. - Save and select.
When to use cloud transcription: mainly when you want Whisper Large quality without the local compute cost. Groq is currently the speed champion. For privacy-sensitive work, stick with local — Parakeet on Apple Silicon is competitive with Whisper Large in real-world use.
Picking the right model — quick guide
- English, Apple Silicon: Parakeet V2.
- English, Intel Mac: Whisper Small or Turbo.
- Other languages, accuracy matters: Whisper Large v3 Turbo.
- Other languages, speed matters: Whisper Small.
- Heavy accent or noisy environment: Whisper Medium minimum, ideally Large.
- Long-form content (lectures, interviews): Whisper Large via Groq cloud — fast and accurate.
Custom dictionary
Independent of the model, the Custom Dictionary lets you correct transcribed text by matching patterns and replacing with your preferred spelling. It runs after transcription but before AI enhancement, so:
- The model can mishear "Ziva" as "Zeeva" — your dictionary catches it.
- Your AI enhancement prompt sees the corrected text, so it doesn't have to guess about your jargon.
Common entries:
- Proper nouns the model gets wrong (your name, your kids' names, your company).
- Technical jargon (TypeScript, Kubernetes, Yjs).
- Spelling preferences (organisation vs. organization).