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

# Languages and translation

> Choose the transcription language in SpeakoFlow, use autodetect, and translate speech to English on your own device. Includes which models cover which languages and which ones can translate.

Two settings control language, and both live with your model in *Settings → Dictation*.

| Setting                  | Default | What it does                                        |
| ------------------------ | ------- | --------------------------------------------------- |
| **Language**             | `Auto`  | Tells the model which language you are speaking.    |
| **Translate to English** | Off     | Transcribes another language straight into English. |

Naming your language instead of leaving it on `Auto` can improve accuracy noticeably, because the model stops spending effort guessing. It helps most on short recordings.

<Note>
  Both rows only appear when your current model supports them. If you cannot find *Translate to English*, the model you are on cannot translate. The whole group is hidden for models with nothing to configure.
</Note>

## Which model for which languages

| Language need                                  | Model                                                      |
| ---------------------------------------------- | ---------------------------------------------------------- |
| Widest coverage, 99 languages                  | **Whisper Medium**, or Whisper Large from the older set    |
| Real-time multilingual, 28 languages           | **Nemotron Streaming 3.5**                                 |
| Highest accuracy across 14 languages           | **Cohere Transcribe**                                      |
| English, German, Spanish, French, instantly    | **Canary 180M Flash**                                      |
| English only, fastest                          | **Parakeet Unified EN 0.6B** (the default)                 |
| 25 mostly-European languages                   | **Parakeet V3** or **Canary 1B v2** (both in Older models) |
| Chinese, English, Japanese, Korean, Cantonese  | **SenseVoice** (Older models)                              |
| Russian                                        | **GigaAM v3** (Older models)                               |
| Taiwanese Mandarin with English code-switching | **Breeze ASR** (Older models)                              |

## Which models translate to English

| Model                                                             | Translates |
| ----------------------------------------------------------------- | ---------- |
| Whisper Small, Medium, Large                                      | Yes        |
| Whisper Turbo                                                     | No         |
| Canary 180M Flash, Canary 1B v2                                   | Yes        |
| Parakeet, Nemotron, Moonshine, Cohere, SenseVoice, GigaAM, Breeze | No         |

Whisper Turbo is the one that catches people out. It is a Whisper model, so translation looks like it should be there, but the model does not support the translation task.

Leave translation off and you get a transcript in the language you spoke. Turn it on and you get English regardless.

<Tip>
  **Non-Western names.** Whisper models are much better at names that trip up English-only models, including Nepali, Indian, and other South Asian names. Add the tricky ones to [Custom words](/docs/dictation/custom-words) so they are spelled your way every time.
</Tip>

<AccordionGroup>
  <Accordion title="Your language choice is only a hint, and can be ignored">
    Your choice is passed to the model **only if that model claims to support the language**. If it does not, SpeakoFlow falls back to autodetection rather than forcing a language the model cannot handle.

    So picking Japanese while an English-only model is active changes nothing.
  </Accordion>

  <Accordion title="Simplified and Traditional Chinese">
    Both are sent to the model as `zh`. The distinction exists in the language list for filtering, not as something the models treat separately.
  </Accordion>

  <Accordion title="Why a model card's capabilities can change after first use">
    Language and translation support are declared by the model file itself. For current-engine models, SpeakoFlow reads those capabilities out of the GGUF the first time the model actually loads and corrects its own catalog entry if it was wrong.
  </Accordion>

  <Accordion title="Custom words work differently per model family">
    Whisper-family models receive them as a recognition prompt, which biases the model before it decides. Every other family gets a fuzzy correction pass over the finished transcript instead.

    Both work. The Whisper route tends to be more accurate because nothing has to be guessed after the fact.
  </Accordion>
</AccordionGroup>