Models
Local Whisper models, your own cloud providers, and OiPer's hosted models.
The Models page is where you download local speech models and register the cloud models your profiles can use. Which of them a dictation actually uses depends on the profile you triggered.
Local Whisper models
Local models are downloaded once and then run entirely on your machine. Larger models are more accurate and slower, and take more disk space and memory.
| Model | Size | Notes |
|---|---|---|
| Tiny | 75 MB | Fastest, least accurate. Fine for short, simple dictation. |
| Base | 142 MB | Good balance for general use. Recommended starting point. |
| Small | 466 MB | Noticeably more accurate than Base. |
| Medium | 1.5 GB | High accuracy, slower. |
| Large v1 / v2 / v3 | 2.9 GB | Highest accuracy, slowest. |
| Large v3 Turbo | 1.5 GB | Large-model accuracy at much lower cost. Recommended. |
Tiny, Base, Small, and Medium also come in English-only variants, which are more accurate than the multilingual model of the same size if you only ever dictate in English.
Models are downloaded from Hugging Face into the models/ folder of OiPer's application
data directory. You can delete a downloaded model from the Models page to reclaim the space.
Hardware backend
Local transcription can run on the CPU or on the GPU. The Whisper Hardware setting in Settings offers:
- Auto — use the GPU when one is usable, otherwise the CPU. The default.
- CPU — always use the CPU. Slower, but avoids GPU driver problems.
- GPU — always use the GPU.
Your own cloud models
You can register cloud models for speech recognition (ASR) and for text formatting (LLM). Each entry stores a provider, a base URL, a model name, and your API key.
Known providers are pre-filled with their base URL and the environment variable they normally use, including OpenAI, Anthropic, OpenRouter, Hugging Face, Groq, Cerebras, Fireworks, Kilo, and OpenCode. Any other OpenAI-compatible endpoint works by entering the base URL yourself.
API keys are kept out of config.json and written to a separate auth.json file in OiPer's
application data directory, so you can share or sync your configuration without leaking
them. Alternatively, point an entry at an environment variable and keep the key outside
OiPer entirely — see Security & Data Handling.
Some models accept audio directly. Marking an entry as accepting audio input lets you transcribe through a general-purpose language model instead of a dedicated speech model, which can help with technical vocabulary.
OiPer models
In oiper mode the models are hosted and maintained for you, and there is no API key to
manage.
Speech models
- Pro — the most accurate transcription, for dictation where every word matters.
- Flash — near real-time speed, trading a little accuracy for the lowest latency.
Language models for formatting, from lightest to most thorough: Auto, Moon, Earth, Mars, Jupiter, and Pluto. Auto adapts to your input and is the low-cost default; Jupiter and Pluto cost more and are meant for demanding rewriting.
Simple mode picks a speech model and a language model for you through a template. Advanced mode lets you choose each one yourself.