# Choose speech models and languages

Compare Apple Speech, Whisper Tiny, Base and Small, Parakeet, and ParakeetV2, then choose language and transcription quality.

Updated: 2026-11-11
Canonical: https://veura.app/guides/speech-models-and-languages

Veura lets you choose the transcription engine that fits your language, storage, speed, and accuracy needs. Downloaded model files remain on the device.

## Compare the models

| Model | Best fit | Displayed download |
| --- | --- | --- |
| Apple Speech | Built into iOS; language availability is managed by the system | System managed |
| Whisper Tiny | Quick dictation and lower device load | About 75 MB |
| Whisper Base | Better accuracy for longer or noisy recordings | 148 MB |
| Whisper Small | Highest Whisper accuracy offered in Veura, with more processing | 487 MB |
| Parakeet | Multilingual Core ML model optimized for Apple chips | About 2.58 GB |
| ParakeetV2 | English-optimized Core ML model with stronger recall | About 2.58 GB |

![Veura model manager showing Apple Speech, Whisper, Parakeet, sizes, and licenses](https://veura.app/screenshots/current/raw/models.png)

The model manager shows download size, license, readiness, and the currently selected model.

## Language behavior

- Apple Speech uses supported languages installed through iOS and does not offer automatic language detection in Veura.
- Whisper and Parakeet can use automatic detection where the model supports it.
- ParakeetV2 is English only and forces English as the transcription language.
- A note records its selected or detected language for later inspection and webhook delivery.

## Choose a quality mode

| Mode | Behavior |
| --- | --- |
| Auto | Uses duration and language to choose speed versus quality |
| Balanced | Faster transcription with lower battery use |
| Balanced+ | Better accuracy at moderate speed |
| Max Quality | Best accuracy, with slower processing and higher battery use |

## Reprocess saved audio

When a note still has saved audio, open Reprocess and choose another ready model or language. Veura keeps the previous text as the original transcript while the new result becomes the working note text. Generated variants are reset when the source transcript changes.

## Model licenses

Whisper model support is shown with the MIT license. Parakeet models are shown with CC BY 4.0 and require attribution. Veura includes third-party notices in Settings; review linked model terms before distributing a modified build.
