Transcription models

The Models page in the sidebar shows four tabs, one per job: Every tab starts with a Runs on switch: “On this computer” or “Cloud”. Switching changes the settings inside the card, not the shape of the page. Below the card is the catalog of models you can download for that job.

​

Speech to text

The catalog has 69 speech models. They are listed in a fixed order, the five recommended ones first, with the models on this computer in their own group above the ones you can download.

​

Recommended models

Pick Parakeet for fast everyday English. Pick Canary 180M Flash on a weak machine (it covers English, German, Spanish, and French). Pick Nemotron for multilingual dictation with automatic language detection. Pick Whisper Medium for the widest language coverage. Pick Cohere when accuracy matters more than speed.

​

The rest of the catalog

Below the recommended five are the other families: more Parakeet and Nemotron variants (including Parakeet TDT 0.6B v2 and v3), every Whisper size from Tiny to Large v3 and Large v3 Turbo, Canary, Cohere Transcribe for Arabic, Granite Speech, Voxtral, Qwen3-ASR, Fun-ASR, Moonshine (with English, Arabic, Chinese, Japanese, Korean, Ukrainian, and Vietnamese variants), SenseVoice Small, GigaAM v3 for Russian, Breeze-ASR-25 for Taiwanese Mandarin, MedASR, and MOSS-Transcribe-Diarize. Each row shows how many languages the model hears. A model with one language says so by name. For more than one, click the count to see the full list. All of them run on the same engine, transcribe.cpp, on your graphics card or processor.

​

Searching and filtering

• Search by name to narrow the list.
• Filter by language with the globe dropdown. It lists every language some model hears, and keeps only the models that hear the one you pick. Simplified and Traditional Chinese count as one language here.
• Streaming keeps only models that transcribe as you speak.
• Translate keeps only models that can translate speech into English.

Filters combine. When nothing matches, Clear filters puts the full list back.

​

Models from an earlier version

Before SpeakoFlow moved every model to transcribe.cpp, some models came in older formats. What happens to them depends on the format.

• Whisper .bin files (Whisper Small, Medium, Turbo, Large, and Breeze ASR) still work. A downloaded one is listed with an Older tag, because the catalog has the same model in the current format. One you never downloaded is not offered any more.
• ONNX models (Parakeet V2 and V3, Moonshine, SenseVoice, GigaAM, Canary, and Cohere) ran on an engine SpeakoFlow no longer includes, so they can’t be used. A downloaded one is listed with a Retired tag and Get new version, which downloads the same model for the new engine. A note above the list gathers them, with one Delete old files button to free the space.

If a retired model was your speech model, it stays selected after the update so you can see what happened. Home and the Speech to text tab show a prompt with Download new version, and pressing a voice shortcut shows “needs an update” on the pill and opens that prompt. When the download finishes, the new version becomes your speech model on its own. Choose another model takes you to the rest of the catalog instead. Until one of those happens, a meeting or a History retranscription tells you the model needs its new version instead of starting.

​

Downloads

Downloads fetch eight chunks at once and resume where they stopped if interrupted. Progress is written to a .partial file, and every retry resumes from there using an HTTP Range request. Each source is retried up to four times with exponential backoff. A server error, a timeout, or a dropped connection is retried. A permanent error like a 404 moves on to the next source. Every catalog model is pinned to one version on Hugging Face, and every file is checked against its SHA-256 checksum when the download finishes. A file that does not match is deleted and the download stops with an error you can retry, so a corrupted or changed file never loads. If Hugging Face fails, the download tries a copy on Handy’s mirror, which is checked the same way. Cancelling keeps the partial file, so resuming later picks up where you left off.

​

Switching models

Pick a different model from the dropdown at the top of the Speech to text tab. You can also switch from the system tray menu without opening the app. A brand-new install is seeded with Parakeet Unified EN 0.6B. If it is not on disk yet, SpeakoFlow falls back to any other downloaded speech model, so the app is never left without a working model.

​

Deleting a model

Click the delete button on any downloaded model to free disk space. You get a confirmation, and a stronger one if it is your active model.

​

The other three tabs

​

AI cleanup

Shows the cleanup provider and model. When no cleanup model is set, cleanup borrows the assistant’s model. The tab has its own catalog of cleanup-suitable models and settings for the cleanup timeout and unload timeout. If AI cleanup is off, the tab shows a button to turn it on. Meeting notes still use this model even when cleanup is off.

​

Assistant

Shows the assistant’s provider and model. The tab has its own catalog of assistant models. If the assistant is turned off, a note says so and links to the Assistant page. AI cleanup and meeting notes fall back to the assistant’s model when they have none of their own.

​

Voice

Shows the voice engine and voice for spoken replies. Covers both local voices (Kokoro, Kitten, Pocket TTS, Supertonic) and cloud voices. See Spoken replies for details.

​

Local vs. cloud per job

Every tab’s Runs on switch lets you move that job between your computer and a cloud provider. Cloud uses a provider’s API and your own key. Your computer is private and free. See Providers for the full list of supported providers and how to set them up.