Speech engines and models

Choose an engine#

Open Settings → Speech recognition. On this device lists the engines that run on your phone, Cloud lists the services that need your own API key. Tap an engine to use it for dictation, the keyboard, the bubble and file transcription. Parakeet is selected after installation.

If the engine still needs its model, its model page opens right away. For engines with a downloadable model, the arrow next to the engine opens the model page at any time. Without Premium, Nemotron Live shows a lock instead.

• Language: Automatic or one fixed language. Parakeet, Nemotron Live and SenseVoice recognize the language by themselves. Moonshine only writes English and GigaAM only Russian. The choice still applies to other engines and to files.
• A profile can use its own engine and language instead of the ones in settings. See Profiles and AI writing.

On-device engines#

These engines work offline once their model is downloaded. Sizes are the one-time download shown in the app.

• Parakeet (on device): Parakeet (about 670 MB) or Parakeet Ultra (about 630 MB), 25 European languages. Fast and accurate, detects the language itself. Ultra was retrained by Moondream and makes fewer errors with the same languages and speed. The best starting point for most people.
• Whisper (on device): 99 languages in five sizes. Whisper Tiny (about 104 MB) is the fastest and least accurate, Whisper Base (about 161 MB) suits short dictations, Whisper Small (about 375 MB) balances speed and accuracy, Whisper Large v3 Turbo (about 1 GB) is the most accurate and takes a few seconds per dictation. Whisper Distil Large v3 (about 984 MB) writes English only and is faster than Turbo. Whisper is slower than Parakeet.
• Canary (on device): Canary 180M Flash (about 207 MB) for English, German, Spanish and French. A third of Parakeet’s size and faster. It needs the dictation language; without one it uses your phone’s language, otherwise English.
• Nemotron Live (on device): shows the text while you speak (about 682 MB). Needs Premium.
• SenseVoice (on device): SenseVoice Small (about 240 MB) for Chinese, Cantonese, English, Japanese and Korean. Detects the language itself and is much faster than Whisper.
• Moonshine (on device): Moonshine Tiny (about 44 MB), English only. The smallest and a fast model, but less accurate than Parakeet and Whisper.
• GigaAM (on device): GigaAM v3 (about 232 MB) for Russian only, with punctuation.
• Android speech recognition: the recognizer built into Android, no download. Needs Android 13 or newer. Which languages work offline depends on the installed speech provider. Use Download selected offline language to install the language you picked.

Download and remove models#

If you dictate with an engine whose model is missing, the Record screen offers Download model first. Large models can use mobile data, so start them on Wi-Fi if your data plan is limited.

1. Open the model page

   In Settings → Speech recognition, tap the arrow next to an on-device engine. Each model shows its size with “stored on this device”, or “Downloaded” once it is on your phone.

2. Tap Download

   The download continues in the background and a notification shows the progress. Without a connection it waits until one is available. Cancel download keeps the downloaded parts, so Resume download continues where it stopped. After a failed download, tap Retry.

3. Pick the model

   Parakeet and Whisper offer several models. Tap one to use it. Each model is downloaded separately.

4. Remove a model you no longer need

   Tap the trash icon next to a downloaded model and confirm with Delete. You can download it again later.

Advanced settings#

Settings → Speech recognition → Advanced holds four options:

• Skip silence: on by default. Trims silence before and after speaking and ignores recordings without speech, so no text gets invented. Nemotron Live receives the audio unfiltered.
• Keep recording after stop: Off, 150 ms (default), 300 ms, 500 ms or 1000 ms, so a last word spoken while tapping stop is not cut off.
• Free model memory: Never, Right after dictating, after 2, 5, 10 (default) or 30 minutes, or After 1 hour. Applies to on-device models; the next dictation loads the model again.
• Fallback engine: None by default. The engine you pick here transcribes instead when the chosen engine fails, for example without network.

Translation, Chinese characters and dictionary terms#

• Translate to English: a switch under Translation on the Whisper and Canary model pages, shared by both engines. Whisper Tiny, Base and Small turn any spoken language into English text, offline; Large v3 Turbo and Distil Large v3 always write what is spoken. Canary translates German, Spanish or French into English. Set Language to the language you speak.
• Chinese characters: in Settings → Text clean-up, choose As recognized, Simplified, Traditional, Traditional (Taiwan) or Traditional (Hong Kong). Chinese dictations are then written in one script; Japanese stays as it is.
• Dictionary terms with Parakeet: the Vocabulary boost (about 132 MB) helps Parakeet recognize your terms. It appears on the Parakeet model page and downloads when you add your first term while Parakeet is selected.
• Dictionary terms with other engines: most cloud engines receive your terms with the recording. Engines that cannot take terms get close misspellings fixed in the text afterwards.

Cloud engines#

Add the API key under Settings → Cloud providers. Each provider page shows Used for, and Test checks the key. Groq, OpenAI, Meta and Own server share one key between speech recognition and AI writing; Own server also shares its Server URL. Keys are stored encrypted on your phone.

• Groq Whisper (cloud): whisper-large-v3-turbo.
• OpenAI (cloud): gpt-transcribe with your dictionary terms.
• ElevenLabs Scribe (cloud): Scribe v2 with your dictionary terms.
• Soniox (cloud): stt-async-v6 with your dictionary terms. Choose the region USA, European Union or Japan on the Soniox page.
• Meta Muse Voice (cloud): muse-voice-transcribe-1.0 with your dictionary terms.
• Reson8 (cloud): TypeWhisper fixes dictionary terms in its text afterwards.
• Deepgram (cloud): Nova-3 with word timings.
• AssemblyAI (cloud): Universal-3.5 Pro with your dictionary terms, many languages.
• Own server: a Whisper-compatible server on your network, such as faster-whisper or LocalAI. Enter Server URL and Speech model under Cloud providers → Own server.