reference
model matrix
Every model yapyap can download — size, languages, what it is for.
Every model below runs on your machine. Sizes are download sizes on disk. The pickers live in Settings → Pipeline and in the setup wizard — this page is the full menu.
live transcription
Shows your words as you speak. Speed matters more than polish — the final transcript is redone by the accurate model afterwards.
| model | size | languages | notes |
|---|---|---|---|
| Whisper Tiny | 75 MB | 99 | fastest |
| Whisper Base | 142 MB | 99 | |
| Whisper Small | 487 MB | 99 | default |
| Whisper Medium | 1.5 GB | 99 |
final transcription
The transcript that everything else builds on.
| model | size | languages | notes |
|---|---|---|---|
| Whisper Medium | 1.5 GB | 99 | |
| Whisper Large v3 Turbo | 1.62 GB | 99 | default — Large-v3 quality, ~6× faster |
| Parakeet TDT v3 | 2.55 GB | 25 European | beats Large v3 on European languages, much faster on CPU |
| Whisper Large v3 | 3.1 GB | 99 | the most thorough |
Whisper picks support a quality setting: Fast (greedy), Balanced, Accurate.
dictation
| model | size | languages | notes |
|---|---|---|---|
| Nemotron 3.5 ASR | 2.6 GB | 35 | punctuation and casing built in |
the supporting cast
Small models the pipeline uses around the transcriber:
| model | size | job |
|---|---|---|
| Silero VAD v5 | 2 MB | hears when speech starts and stops |
| pyannote (community) | 6 MB | cuts the audio into speaker turns |
| WeSpeaker ResNet34 | 27 MB | tells the voices apart |
| GLiNER multi v2.1 | 211 MB | finds names, dates, places |
| XLM-RoBERTa sentiment | 1.1 GB | reads the mood of a passage |
| Fullstop (47 languages) | 1.1 GB | punctuation — default |
| MiniLM-L12 multilingual | 470 MB | powers search — default |
| MiniLM-L6 English | 91 MB | search, English only, lighter |
language models
The models that write titles and run lenses are a separate story with their own engines and bring-your-own options — see language model engines.
| model | notes |
|---|---|
| Qwen 2.5 7B | default |
| Qwen 2.5 3B | lighter |
| Qwen 2.5 14B | heavier, better on long meetings |
| Phi 3.5 mini | no tool use — a lens that calls tools will skip |
| Llama 3.2 3B | restrictive license |
| Gemma 2 2B | restrictive license, no tool use |
notes
- Models are downloaded once and shared between stages. yapyap detects models you already have and does not download them twice.
- One model per stage stays warm in memory and is evicted when idle. Only the language model gets a RAM advisory when it is large for your machine.