> ## Documentation Index
> Fetch the complete documentation index at: https://docs.stenoai.co/llms.txt
> Use this file to discover all available pages before exploring further.

# Transcription models

> Steno uses Whisper models for local transcription. Choose the right model based on your accuracy requirements and available hardware.

Steno transcribes recordings using [whisper.cpp](https://github.com/ggerganov/whisper.cpp), a local port of OpenAI's Whisper model. All transcription runs on your Mac -- no audio is sent to any server.

## Available models

| Model            | Size    | Speed    | Accuracy  | Best for                              |
| ---------------- | ------- | -------- | --------- | ------------------------------------- |
| `whisper-base`   | \~150MB | Fastest  | Good      | Quick notes, clear audio              |
| `whisper-small`  | \~480MB | Fast     | Very good | Most meetings **(default)**           |
| `whisper-medium` | \~1.5GB | Moderate | Excellent | Important recordings, difficult audio |

`whisper-small` is the default and handles most meetings well. The larger models take longer to process but produce noticeably better results for noisy recordings, strong accents, or technical terminology.

## How to change models

1. Open **Settings** in Steno
2. Select **Models → Transcription**
3. Choose a model from the list
4. If the model is not yet downloaded, Steno will prompt you to download it

The model is applied to all future recordings. Existing recordings are not re-transcribed automatically, but you can re-run transcription on any recording from its detail view.

## Speed estimates (Apple Silicon)

Transcription runs faster than real-time on Apple Silicon. These are approximate figures for a 30-minute recording on an M-series Mac:

| Model            | Processing time |
| ---------------- | --------------- |
| `whisper-base`   | \~1-2 minutes   |
| `whisper-small`  | \~2-4 minutes   |
| `whisper-medium` | \~5-10 minutes  |

On Intel Macs, expect roughly 2-3× slower processing times.

## Language support

All Whisper models support 99 languages. Steno auto-detects the language from the first few seconds of audio -- you do not need to configure this manually. English, Spanish, French, German, Japanese, Chinese, and Portuguese perform especially well.

## Accuracy tips

* **Record in a quiet environment** -- background noise is the biggest accuracy factor
* **Use a good microphone** -- system audio capture (both call sides) is more accurate than microphone-only for virtual meetings
* **Use `whisper-medium` for technical content** -- it handles domain-specific terminology, proper nouns, and acronyms better than the smaller models

***

<Accordion title="Which model should I start with?">
  Start with `whisper-small` (the default). It is fast and accurate for most meeting types. Switch to `whisper-medium` if you notice accuracy issues with important recordings.
</Accordion>

<Accordion title="Can I re-transcribe a recording with a different model?">
  Yes. Open the recording, click the options menu, and select **Re-transcribe**. Choose the model you want to use. This overwrites the existing transcript.
</Accordion>

<Accordion title="Does transcription accuracy vary by language?">
  Yes. Whisper was trained on significantly more English data than other languages. English, Spanish, French, and German accuracy is high. Less common languages may have lower accuracy, particularly for technical vocabulary.
</Accordion>