> ## Documentation Index
> Fetch the complete documentation index at: https://hyperwhisper.com/docs/llms.txt
> Use this file to discover all available pages before exploring further.
# Language & Detection
> Configure the transcription language per mode, understand automatic language detection, and learn how language choice affects engine availability, vocabulary boosting, and AI post-processing.
HyperWhisper lets you set a transcription language for each mode independently. You can also let the engine detect the language automatically from the audio, though this works best with longer recordings.
## Setting the Language for a Mode
Language is stored **per mode**, so each mode in your list can target a different language. This makes it easy to keep a dedicated mode for each language you use regularly without changing any setting between tasks.
Click the HyperWhisper menu bar icon to open the main window, then click **Modes** in the sidebar. Click the mode you want to edit, or click **Create Mode** to create one.
In the mode editor, find the **Language** row under the Transcription section. Click the dropdown to open the full language picker. Popular languages appear at the top; the rest are listed alphabetically below them.
Click **Save**. The language choice is stored with the mode and takes effect on the next recording in that mode.
In the main window sidebar, click **Modes**, then click the edit icon next to the mode you want to change, or click **Create Mode**.
Scroll to the **Language** section in the mode editor. Click the dropdown to choose from the available languages for the selected model.
Click **Save**. The language is stored with that mode.
On iOS, HyperWhisper transcribes through HyperWhisper Cloud. The per-mode **Language** section in the mode editor controls the **English Spelling** variant used by the AI post-processor (American, British, Australian, or Canadian). The transcription language itself is sent as part of the mode's locale when your recording is uploaded.
The default selection is **Automatic**, which lets the engine detect the language from the audio. See [Automatic Language Detection](#automatic-language-detection) below for when to use it and its limitations.
Create a separate mode for each language you use regularly. Switching modes is instant via the mode picker or keyboard shortcut, and each mode remembers its own language — no manual changes needed between sessions.
***
## Automatic Language Detection
When you select **Automatic**, HyperWhisper does not send a language hint to the transcription engine. The engine infers the language directly from the audio content.
**Auto-detect works best when:**
* Your recording contains at least 10–15 seconds of clear speech
* You are dictating in a single language throughout the recording
* The language you are speaking is well-represented in the model's training data
**Auto-detect is unreliable when:**
* Recordings are short (under 10–15 seconds) — there may not be enough speech for the engine to identify the language with confidence
* You switch languages mid-recording
Short auto-detect recordings can produce nonsense output, empty results, or the wrong language. For most use cases, setting an explicit language gives better and more consistent results. See [Best Practices](/best-practices) for more detail.
Some models and cloud providers do not support auto-detect. When a model only supports English, the language picker is hidden entirely in the mode editor and the engine always runs in English-only mode.
***
## English-Only Models
Several local models support only English. For these models, the language picker is hidden in the mode editor and language detection is not available:
* **Whisper `.en` variants** (e.g., `base.en`, `small.en`) — English-only builds of Whisper, optimized for English accuracy
* **Parakeet V2** — NVIDIA's English-only Parakeet model with the highest recall
When you select one of these models, the language field shows an informational notice instead of the picker, and the mode will always transcribe in English.
***
## Supported Languages by Model
The number of available languages depends on the model or provider you choose.
| Model / Provider | Languages |
| ----------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Whisper (all multilingual variants)** | 101 languages |
| **HyperWhisper Cloud, OpenAI, Groq, Deepgram, AssemblyAI, ElevenLabs, Soniox, Mistral, Gemini, Grok** | Varies by model — picker filters automatically |
| **Google Speech / Chirp 3** (via HyperWhisper Cloud) | \~88 languages |
| **Azure MAI-Transcribe** (via HyperWhisper Cloud) | \~42 languages |
| **Parakeet V3** | 25 European languages (English, German, French, Spanish, Italian, Portuguese, Dutch, Polish, Russian, Ukrainian, Czech, Slovak, Hungarian, Romanian, Bulgarian, Croatian, Slovenian, Serbian, Danish, Swedish, Norwegian, Finnish, Estonian, Latvian, Lithuanian) |
| **Qwen3 ASR** *(macOS only)* | 30 languages including Chinese, English, Cantonese, Arabic, Japanese, Korean, and major European languages |
| **Parakeet V2** | English only (picker hidden) |
| **Whisper `.en` variants** | English only (picker hidden) |
For cloud providers, the language picker filters itself to only show the languages the selected model actually supports. Switching to a model with a narrower language list will automatically fall back to the first available language if your current selection is not supported by the new model.
See the [Model Library](/models) for the full per-model breakdown.
**Popular languages appear at the top of the picker** on both platforms: English, Japanese, Spanish, Chinese, Chinese (Traditional), Dutch, Hindi, Russian, Korean, Italian, Ukrainian, Polish, Portuguese, Greek, Czech, Swedish, Norwegian, Danish, and Indonesian. All other languages follow in alphabetical order.
***
## How Language Affects Other Features
Your language choice in a mode affects more than just the transcription engine.
### Vocabulary Boosting
Custom vocabulary words you add in [Vocabulary](/vocabulary) are sent to cloud providers as hints to improve accuracy on proper nouns, technical terms, and domain-specific words.
**Deepgram Nova-3 specifically:** vocabulary boosting via the `keyterm` parameter is ignored when the language is set to **Automatic**. To get vocabulary boosting with Deepgram Nova-3, set an explicit language in the mode. When you select Nova-3 with auto-detect enabled, a notice appears in the mode editor.
Some providers do not support vocabulary hints at all (the mode editor shows a notice in those cases).
### AI Post-Processing
When AI post-processing is enabled, the transcription result is refined by a language model. The **English Spelling** setting (American, British, Australian, or Canadian) is only available — and only meaningful — when the mode language is set to English. For non-English modes, spelling conventions follow the target language automatically.
### Engine Availability
Choosing a language that a particular model does not support will cause the picker to fall back to the first language in the model's supported list. If you switch models while on a language the new model does not support, the language resets automatically rather than sending an unsupported code to the engine.
***
## Streaming Language
Streaming transcription has its **own language setting**, separate from per-mode language.
The streaming language is set in the **Streaming** section of the sidebar. It defaults to **English**. You can change it to any supported language or to **Automatic**. The available options filter to match the streaming provider and model you have selected — for example, Parakeet V3 streaming limits the picker to its 25 supported languages.
Open the **Streaming** page in the main sidebar. The **Language** dropdown shows all supported languages for the selected streaming provider. It defaults to **English**.
The same auto-detect caveats apply to streaming: short or single-sentence inputs are likely to produce inconsistent results when language is set to **Automatic**.
The Deepgram vocabulary boosting limitation applies to streaming as well: setting the streaming language to **Automatic** disables vocabulary hints for Deepgram Nova-3.