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

# Models & Engines

> Choose the right speech-to-text model for your use case.

Vowen runs your dictation through a speech model of your choice. Some run entirely on your machine, some are cloud APIs you connect with your own key, and one lets you point Vowen at a server you run yourself.

All of them live on the **Speech models** page in the left sidebar. That page has three sections: **Cloud**, **Local** (split into English Only and Multilingual), and, on Windows, **GPU Acceleration**.

<Frame caption="The Speech models page. Cloud providers come first, then the Local sections. The dark card with a tick is your active model.">
  <img src="https://mintcdn.com/vowen/94qaixcfbwWHsixE/images/app/models-page-overview.png?fit=max&auto=format&n=94qaixcfbwWHsixE&q=85&s=7cfe8aa5ddf876da432b3061de490e3c" alt="Vowen's Speech models page showing a grid of cloud provider cards above the Local section" width="2400" height="1600" data-path="images/app/models-page-overview.png" />
</Frame>

## Quick recommendations

| Use case | Recommended model | Why |
| - | - | - |
| Everyday English (macOS, Windows) | Parakeet V2 | The default. Offline, fast, English-only |
| Everyday multilingual (macOS, Windows) | Parakeet V3 | Offline, auto-detects 25 European languages |
| Live preview with punctuation (macOS) | Nemotron EN or Nemotron Multilingual | The only local models whose partial results are already punctuated |
| Maximum offline accuracy | Whisper Large v3 Turbo | Best local accuracy per second of processing |
| Fastest cloud accuracy on a free tier | Whisper Large v3 Turbo (Groq) | Generous free tier, very fast |
| Non-English via cloud | Whisper Large v3 (Groq) or Nova 3 (Deepgram) | Strong multilingual coverage |
| Indian languages | Saaras v3 (Sarvam AI) | 22 Indian languages plus English |
| Japanese | Parakeet Japanese (macOS) or a cloud model | Purpose-trained local option |
| Mandarin | Parakeet Mandarin (macOS) or a cloud model | Purpose-trained local option |
| Meetings where you need speaker labels | Nova 3, Scribe v2, Universal-3.5 Pro, or local Parakeet / Whisper | See [Speaker identification](#speaker-identification) |
| Maximum privacy | Any local model | Audio never leaves your device |

## Local models

Local models run entirely on your machine and need no internet connection. Nothing is bundled with the installer. You download every local model on demand from the Speech models page. See [Model Download & Management](/transcription/model-download).

Sizes below are the download sizes as the app reports them.

### Parakeet

NVIDIA's Parakeet family. Vowen ships four separate Parakeet models, not one. Each is a different set of weights with its own language coverage and its own capabilities.

| Model | Platforms | Download | Languages | Live streaming | Timestamps | Speaker labels |
| - | - | - | - | - | - | - |
| **Parakeet V2** (default) | macOS, Windows | 430 MB (macOS) / 451 MB (Windows) | English only | Yes (macOS) | Yes | Yes |
| **Parakeet V3** | macOS, Windows | 445 MB (macOS) / 478 MB (Windows) | 25 European languages, auto-detected | Yes (macOS) | Yes | Yes |
| **Parakeet Japanese** | macOS | 532 MB | Japanese only | No | Yes | No |
| **Parakeet Mandarin** | macOS | 1.5 GB | Mandarin Chinese only | No | No | No |

<Note>
  **Parakeet Mandarin** returns plain text only: no per-word timings, so no clickable timestamps and no subtitle export, no live preview, and no speaker labels. Everything else about it works normally.
</Note>

<Note>
  **Parakeet Japanese and Mandarin are macOS only.** There is no Windows build, and on Windows the Speech models page simply does not list them.
</Note>

On macOS, Parakeet V2 and V3 also pull down a small **end-of-utterance streaming model** in the background. It is what powers the live transcription preview. You never pick it or see it in the model list. See [the download page](/transcription/model-download#the-hidden-streaming-model) for what it is and when it gets removed.

### Nemotron (macOS, Apple Silicon)

NVIDIA Nemotron Speech models. Both stream natively, and unlike Parakeet their live preview arrives already punctuated and capitalized. It reads as finished text while you are still speaking, instead of arriving as a lowercase run-on that gets fixed at the end.

| Model | Platforms | Download | Languages | Live streaming | Timestamps | Speaker labels |
| - | - | - | - | - | - | - |
| **Nemotron EN** | macOS (Apple Silicon) | 528 MB | English only | Yes, punctuated | No | No |
| **Nemotron Multilingual** | macOS (Apple Silicon) | 588 MB | 31 languages | Yes, punctuated | No | No |

Nemotron transcribes files with the same engine it uses for dictation, so its output carries no per-segment timings.

See [Languages](/transcription/languages#nemotron-multilingual) for the full 31-language list.

### Whisper

OpenAI's Whisper, running locally on your own machine. 99 languages on the multilingual variants.

| Model | Download | Speed | Accuracy | Best for |
| - | - | - | - | - |
| Tiny | 74 MB | Fastest | Basic | Quick notes, testing |
| Tiny (EN) | 74 MB | Fastest | Basic (English) | Fast English dictation |
| Base | 141 MB | Fast | Good | Multilingual basics |
| Base (EN) | 141 MB | Fast | Good | General English use |
| Small | 465 MB | Medium | Great | Professional work |
| Small (EN) | 465 MB | Medium | Great (English) | Detailed English |
| Medium | 1.4 GB | Slow | Excellent | High-quality output |
| Medium (EN) | 1.4 GB | Slow | Excellent (English) | Long-form English |
| Large v3 | 2.9 GB | Slowest | Best | Maximum accuracy |
| Large v3 Turbo | 1.5 GB | Medium-Fast | Excellent | Best balance |

<Note>
  The **(EN)** models are English-only and slightly more accurate for English than their multilingual counterparts of the same size.
</Note>

<Warning>
  **Tiny and Base ignore your Dictionary.** Among the local Whisper sizes, only Small, Medium, Large v3 and Large v3 Turbo use your vocabulary. If you rely on custom vocabulary, use Small or larger. See [Accuracy](/transcription/accuracy#how-vocabulary-actually-reaches-each-model).
</Warning>

### Start-up cost

Ordinary dictations pay no model load cost. On macOS and Windows alike, a local model is loaded once and stays ready between dictations.

You do pay a one-time wait the first time a given model runs on your Mac, while it is optimized for your specific chip. That takes roughly 20 seconds, and noticeably longer the very first time. During that window the indicator shows "Optimizing Parakeet model", and a dictation you start anyway is saved for retry rather than transcribed. See [the download page](/transcription/model-download#first-run-optimization) for what to expect.

## Cloud models

Cloud models send audio to a third-party API. They need an internet connection and an API key that you supply. Vowen never charges you for cloud transcription; you pay the provider directly, and most have a free tier.

Counts below are the languages Vowen offers in the dictation language picker for that model. See [Languages](/transcription/languages#the-language-list-changes-with-your-model) for the lists themselves.

| Model | Provider | Dictation languages | Real-time streaming |
| - | - | - | - |
| Whisper Large v3 Turbo | Groq | 99 | No |
| Whisper Large v3 | Groq | 99 | No |
| Nova 2 | Deepgram | 33 | Yes |
| Nova 3 | Deepgram | 50 | Yes |
| Scribe v2 | ElevenLabs | Multilingual | Yes |
| Universal-3.5 Pro | AssemblyAI | 6 streaming / multilingual batch | Yes |
| Voxtral Mini | Mistral | 13 | Yes |
| Saaras v3 | Sarvam AI | 23 (22 Indian + English) | Yes |
| Real-Time STT | Soniox | 60+ | Yes |
| Aurora | xAI | 25 | Yes |
| Ink 2 | Cartesia | English | Yes |
| Speechmatics | Speechmatics | 39 | Yes |
| OpenAI | OpenAI | Multilingual | Yes |
| Gemini (Live + Flash) | Google | Multilingual | Yes |
| Custom Server | Yours | Depends on your server | No |

<AccordionGroup>
  <Accordion title="OpenAI runs gpt-live-transcribe and gpt-transcribe">
    There is no `gpt-4o-transcribe` in Vowen. Dictation uses either **`gpt-live-transcribe`** (streams while you speak) or **`gpt-transcribe`** (transcribes after you stop), depending on the toggle described in [Per-model options](#per-model-options). File transcription always uses `gpt-transcribe`, except when speaker identification is on, which switches to `gpt-4o-transcribe-diarize`. That is the only OpenAI model that emits speaker labels, and the only one that accepts no vocabulary hints at all.
  </Accordion>

  <Accordion title="AssemblyAI ships Universal-3.5 Pro">
    Dictation streams over WebSocket in six languages: English, Spanish, French, German, Italian and Portuguese. File transcription uses Universal-3.5 Pro for the 18 languages it natively supports and falls back to Universal-2 for anything else.
  </Accordion>

  <Accordion title="Gemini has two engines behind one entry">
    Streaming uses Gemini Live, whose sessions are capped at 15 minutes each. File transcription uses Gemini Flash and handles audio up to about 9.5 hours. Custom vocabulary and speaker labels both work on Gemini file transcription, but more weakly than on providers built for them.
  </Accordion>

  <Accordion title="Soniox">
    Listed in the app as **Real-Time STT (Soniox)**. One entry covers both a real-time streaming model for dictation and an async model for file transcription, across 60+ languages. It is the strongest option in Vowen for switching languages mid-sentence.
  </Accordion>

  <Accordion title="Cartesia Ink 2">
    Streaming-first, tuned for the lowest word error rate and the fastest live preview. **English only.** File transcription runs Cartesia's batch model, which does not use your Dictionary.
  </Accordion>

  <Accordion title="Groq has no native diarization">
    Groq's API returns text only. When you turn on speaker identification with a Groq model, Vowen runs its own on-device speaker pipeline over the audio afterwards, on macOS and Windows. See [Diarization](/meeting-notes/diarization).
  </Accordion>
</AccordionGroup>

### Setting up a cloud model

<Steps>
  <Step title="Open Speech models">
    Click **Speech models** in the left sidebar. Cloud models are the first section.
  </Step>

  <Step title="Pick a provider">
    Click its card. A configuration panel opens showing the model's languages and capabilities.

    This is true only for a provider you have not set up yet. Once a provider is configured, clicking its card **switches to that model** instead. Use the gear on the card to reopen its settings.

    <Frame caption="Click any card in the Cloud section to open its configuration panel.">
      <img src="https://mintcdn.com/vowen/94qaixcfbwWHsixE/images/app/models-cloud-setup-1.png?fit=max&auto=format&n=94qaixcfbwWHsixE&q=85&s=9ef62373ca57ca46f79d0250a4e588d7" alt="The Cloud section of the Speech models page with one provider card highlighted" width="2400" height="1600" data-path="images/app/models-cloud-setup-1.png" />
    </Frame>
  </Step>

  <Step title="Paste an API key">
    Use the "Get your API key here" link to open the provider's dashboard, generate a key, and paste it in.

    <Frame caption="The panel opens inline, underneath the card. Paste the key and click Save.">
      <img src="https://mintcdn.com/vowen/94qaixcfbwWHsixE/images/app/models-cloud-setup-2.png?fit=max&auto=format&n=94qaixcfbwWHsixE&q=85&s=4954c6705253d0a9e08bdfd990b26a28" alt="The Speech models page with a provider's configuration panel open inline, showing an API Key field with Cancel, Save and Disconnect buttons" width="2400" height="1600" data-path="images/app/models-cloud-setup-2.png" />
    </Frame>
  </Step>

  <Step title="Save">
    The model becomes your active model immediately. You can connect as many providers as you like and switch between them at any time, including from the tray menu.

    <Frame caption="A provider you have set up keeps a gear on its card. The gear reopens the key panel; clicking the card itself switches to that model.">
      <img src="https://mintcdn.com/vowen/94qaixcfbwWHsixE/images/app/models-cloud-setup-3.png?fit=max&auto=format&n=94qaixcfbwWHsixE&q=85&s=34405aed78b66181ed025534035f57f8" alt="The Cloud section with a configured provider card highlighted, showing the gear that reopens its settings" width="2400" height="1600" data-path="images/app/models-cloud-setup-3.png" />
    </Frame>
  </Step>
</Steps>

<Tip>Groq is the most popular choice among Vowen users. It is fast, accurate, and its free tier covers most daily use.</Tip>

### Per-model options

Two providers expose extra controls in the same panel where you paste the API key.

For OpenAI, **Dictation model** is a two-way toggle:

| Option | Behavior |
| - | - |
| `gpt-live-transcribe` | Streams while you speak. Fastest results. Supports custom vocabulary. |
| `gpt-transcribe` | Transcribes after you stop speaking. Slightly slower, cheaper per minute. Supports custom vocabulary. |

This affects dictation only. File transcription always uses `gpt-transcribe`.

<Frame caption="OpenAI's Dictation model toggle, in the same panel as the API key.">
  <img src="https://mintcdn.com/vowen/94qaixcfbwWHsixE/images/app/models-openai-dictation-mode.png?fit=max&auto=format&n=94qaixcfbwWHsixE&q=85&s=ffb759a03e649a9782e8c87a6fb08698" alt="The OpenAI configuration panel with the Dictation model toggle highlighted, offering gpt-live-transcribe and gpt-transcribe" width="2400" height="1600" data-path="images/app/models-openai-dictation-mode.png" />
</Frame>

**ElevenLabs Scribe v2** has two toggles, both off by default:

| Toggle | What it does |
| - | - |
| Remove filler words and false starts | ElevenLabs cleans the transcript itself before returning it |
| Tag audio events like laughter | Non-speech events appear in the transcript |

<Frame caption="ElevenLabs adds its two toggles below the key field. Both are off unless you turn them on.">
  <img src="https://mintcdn.com/vowen/94qaixcfbwWHsixE/images/app/models-elevenlabs-toggles.png?fit=max&auto=format&n=94qaixcfbwWHsixE&q=85&s=2387eb504b5370115e8b15d2c3e5b75f" alt="The ElevenLabs Scribe v2 configuration panel with its two option toggles highlighted" width="2400" height="1600" data-path="images/app/models-elevenlabs-toggles.png" />
</Frame>

## Custom speech servers

Point Vowen at any **OpenAI-compatible transcription endpoint**: a local whisper.cpp server, LocalAI, Ollama, FastFlowLM, a vLLM Whisper deployment, or a hosted service that speaks the same API.

A custom server needs a **Server URL**. A **Model Name** and an **API Key** are both optional; leave them blank when your server does not require them. Vowen checks that the URL is reachable before saving, so a typo fails immediately instead of surfacing at your next dictation.

Once saved, a custom server behaves like any other cloud model: it renders as its own card on the Speech models page, appears in every model picker, and can be selected from the tray.

<Frame caption="The form: a name, the server URL, and optionally a model name and key.">
  <img src="https://mintcdn.com/vowen/94qaixcfbwWHsixE/images/app/models-custom-server-2.png?fit=max&auto=format&n=94qaixcfbwWHsixE&q=85&s=46684ef349cc7a7ba34fe6d72b785ca7" alt="The Add Custom API form filled in with a name, a server URL, a model name, and an empty optional API key field" width="2400" height="1600" data-path="images/app/models-custom-server-2.png" />
</Frame>

<Frame caption="Add Custom API sits at the end of the Cloud section, beside the built-in Local / Self-hosted card.">
  <img src="https://mintcdn.com/vowen/94qaixcfbwWHsixE/images/app/models-custom-server-1.png?fit=max&auto=format&n=94qaixcfbwWHsixE&q=85&s=96d33e68c4d950ada4118e61f813224d" alt="The end of the Cloud section on the Speech models page, with the dotted Add Custom API card highlighted" width="2400" height="1600" data-path="images/app/models-custom-server-1.png" />
</Frame>

| Plan | Custom servers |
| - | - |
| Free | One server, using the built-in **Custom Server** card |
| Pro | Unlimited servers, each with a name you choose |

<Note>
  Adding a **second and further** custom server is a <a href="https://vowen.ai" className="font-semibold underline-offset-2" style={{ color: "#8b5cf6", backgroundColor: "rgba(139, 92, 246, 0.12)", border: "1px solid rgba(139, 92, 246, 0.35)", padding: "2px 8px", borderRadius: "6px", fontSize: "0.85em", textDecoration: "none", whiteSpace: "nowrap" }}>Pro</a> feature. Use **Add Custom API** at the end of the cloud section. Free users keep the single original slot.
</Note>

Custom servers receive your Dictionary terms the same way Groq does, as a hint. It helps, but less than an engine that uses your Dictionary directly.

## Speaker identification

Not every model can label who said what. These can, for pre-recorded audio and meeting notes:

**Cloud:** Nova 2 and Nova 3 (Deepgram), Scribe v2 (ElevenLabs), Universal-3.5 Pro (AssemblyAI), Voxtral Mini (Mistral), Aurora (xAI), Speechmatics, OpenAI, Gemini.

**Local, on macOS and Windows:** Parakeet V2, Parakeet V3, and every Whisper model. Groq's two Whisper models also qualify, because Vowen runs the same on-device speaker pipeline over their output.

Nemotron, Parakeet Japanese, Parakeet Mandarin, Cartesia Ink 2, Saaras v3 and custom servers do not offer speaker labels.

Speaker identification is a <a href="https://vowen.ai" className="font-semibold underline-offset-2" style={{ color: "#8b5cf6", backgroundColor: "rgba(139, 92, 246, 0.12)", border: "1px solid rgba(139, 92, 246, 0.35)", padding: "2px 8px", borderRadius: "6px", fontSize: "0.85em", textDecoration: "none", whiteSpace: "nowrap" }}>Pro</a> feature. Full detail in [Diarization](/meeting-notes/diarization).

## Real-time preview

Live text under the recording indicator needs a streaming-capable model. Every cloud model in the table above except the two Groq entries and custom servers can stream. Locally, Parakeet V2, Parakeet V3, Nemotron EN and Nemotron Multilingual stream on macOS; Parakeet streaming is not available on Windows.

Turn it on in **Settings > Recording > Real-time transcription preview**. The row is greyed out until your active model can stream.

## GPU acceleration (Windows)

If you have an NVIDIA GPU, Vowen can download a CUDA acceleration module that dramatically speeds up local Whisper transcription.

* **Windows only.** macOS is already accelerated and needs nothing extra; Linux is not supported.
* Requires an **RTX 2000 series or newer** NVIDIA GPU.
* The download is about **631 MB**. No separate CUDA toolkit install is needed.

The card only appears on the Speech models page when Vowen detects a supported GPU. Download it, restart Vowen, and even Large v3 responds in a couple of seconds.

## Local or cloud

| Factor | Local | Cloud |
| - | - | - |
| Privacy | Audio never leaves your device | Audio is sent to the provider |
| Speed (macOS) | Very fast with Parakeet or Nemotron | Fast |
| Speed (Windows) | Fast with Parakeet; slow for large Whisper without a GPU | Fast |
| Accuracy | Good to excellent | Excellent |
| Internet | Not required | Required |
| Cost | Free | Free tier or paid API |
| Custom vocabulary | Whisper (Small and up) only | Used well by most providers |
| Languages | 99 (Whisper), 31 (Nemotron), 25 (Parakeet V3) | Varies by provider |


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.