> ## 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.
# Transcription models
> The full catalog of speech-to-text models SpeakoFlow can download and run locally, with size, language coverage, streaming, and translation for each. Also covers older models, downloads, and deleting models.
Every model runs on your machine. They trade accuracy for speed, so pick the one that fits your hardware and languages.
Open the catalog from *Settings → Dictation → Change model*.
## Recommended models
These five run on SpeakoFlow's current engine, a native ggml/GGUF build. Sizes are the exact download.
| Model | Size | Languages | Streams | Translates |
| -------------------------------------- | ------ | ------------ | ------- | ---------- |
| **Parakeet Unified EN 0.6B** (default) | 697 MB | English only | Yes | No |
| **Nemotron Streaming 3.5** | 716 MB | 28 | Yes | No |
| **Canary 180M Flash** | 208 MB | 4 | No | Yes |
| **Cohere Transcribe** | 1.7 GB | 14 | No | No |
| **Whisper Medium** | 793 MB | 99 | No | Yes |
Pick **Parakeet Unified EN 0.6B** for fast everyday English, **Canary 180M Flash** on a weak machine, **Whisper Medium** for the widest language coverage, and **Cohere Transcribe** when accuracy matters more than speed. Canary covers English, German, Spanish, and French.
Only catalog entries flagged `recommended` appear in this list. Everything else lives under **Older models**.
Streaming models can transcribe as you speak, which is what [live transcription](/docs/dictation/live-transcription) needs. Everything else transcribes in one pass when you stop talking.
## Managing models
The catalog splits into **Downloaded models**, **Available to download**, and **Older models**. Your active model sorts first, then the recommended order.
* **Search by name** to filter the list. Searching also reveals the older models.
* **Filter by language** with the globe dropdown. It lists only models that claim support for the language you pick.
* **Delete** a model to free disk space. You get a confirmation, and a stronger one if it is your active model.
* **Switch from the tray.** The system tray menu has a model submenu, so you can change models without opening Settings.
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 transcription model you have downloaded, so the app is never stranded without a working model.
These run on the previous engine (ONNX and whisper.cpp via transcribe-rs) rather than the current one. They are still fully supported and were never removed. They are just no longer the first suggestion.
* **Whisper Small, Medium, Turbo, Large.** Nearly 100 languages. Small, Medium, and Large translate to English. Turbo does not.
* **Parakeet V2** (English only) and **Parakeet V3** (25 languages, mostly European, plus Russian and Ukrainian).
* **Moonshine Base** and **Moonshine V2 Tiny, Small, Medium.** English only, extremely fast, good with accents. The V2 variants stream.
* **Canary 180M Flash** (English, German, Spanish, French) and **Canary 1B v2** (25 European languages). Both translate to English.
* **Cohere.** Large and slower, very accurate across 14 languages.
* **SenseVoice.** Chinese, English, Japanese, Korean, and Cantonese. Very fast.
* **GigaAM v3.** Russian only.
* **Breeze ASR.** Tuned for Taiwanese Mandarin with English code-switching.
The group stays folded so a first-timer only sees the recommended few. It opens automatically as soon as you type in the search box.
Progress is written to a `.partial` file next to the model, and every retry resumes from where it stopped using an HTTP `Range` request. If the server ignores the range and sends the whole file again, SpeakoFlow notices and restarts cleanly rather than appending garbage.
Each source is retried up to four times with exponential backoff. A server error, a timeout, or a dropped connection is treated as worth retrying. A permanent client error such as a 404 is not, and moves straight on.
Cancelling keeps the partial file, so resuming later picks up where you left off.
Older-engine models ship with a SHA-256 checksum and are verified after download.
The current GGUF catalog carries no per-file hashes, so those models are **not** checksum-verified.
If a download looks stuck at 0%, it is usually still negotiating the redirect to Hugging Face's CDN. Give it a few seconds before cancelling.
Which model to pick for which languages.
GPU acceleration, idle unloading, and context size.