WinSTTModels
Browse the full WinSTT speech-to-text catalog — 70+ on-device ONNX models across 12 families, with accuracy, speed, size, and language coverage so you can pick the right trade-off.
WinSTT ships 70+ speech-to-text models across 12 families, all running on a single ONNX engine on your own hardware. Pick one in the Model tab and the app downloads it in the background — nothing about your audio leaves the machine.
One engine, many models
Every model — Whisper included — runs through the same inference engine (ONNX Runtime via
the onnx-asr library). There is no separate backend to choose: switching models is a
download, not a re-install. Most swaps happen live without restarting the server.
Whether inference runs on the GPU or CPU is decided by the installer you chose and the Device setting, not by the model. Numerical precision is chosen per-download via quantization.
The 12 families
Each row sums up one family: how many models it offers, the typical on-disk size at default precision, the languages it covers, and where it fits best. Lower WER (word error rate) is more accurate; higher RTFx (real-time factor) is faster.
| Family | Models | Typical size | Languages | Recommended for |
|---|---|---|---|---|
| Whisper | 12 | 149 MB – 4.5 GB | 99 (multilingual + EN-only) | The all-rounder — accuracy at every size, broadest language reach |
| Lite-Whisper | 3 | ~1.1 GB | 99 | Compressed Whisper Large v3 Turbo — large-model quality at smaller size |
| NeMo | 34 | 0.1 – 4 GB | English, Russian, and 25-lang multilingual | Fast offline Parakeet / Canary plus native streaming exports |
| Moonshine | 10 | 0.1 – 0.7 GB | EN + per-language fine-tunes | Lowest-latency real-time on CPU; tiny footprint |
| Cohere | 1 | ~5.1 GB | 14 | Top-tier accuracy — near the top of the HF Open ASR leaderboard |
| Granite Speech | 2 | 2.5 – 16 GB | en · fr · de · es · pt | IBM autoregressive and non-autoregressive speech models with strong punctuation |
| Qwen3-ASR | 2 | 2.0 – 4.1 GB | 30 | Alibaba Qwen LLM-decoder ASR with auto language detection; int4-only downloads |
| Audio8-ASR / ARK-ASR | 3 | 0.67 – 4.4 GB | 7 · 19 | LLM-decoder ASR at three sizes — the 0.1B runs on any CPU (non-commercial licence); the Apache-2.0 ARK-ASR pair tops the accuracy table at 3B |
| SenseVoice | 1 | 159 MB | 5 (zh/en/ja/ko/yue) | Fast East-Asian multilingual in a tiny int8 graph |
| Dolphin | 1 | 104 MB | 15 Asian and Middle Eastern languages | Compact multilingual CTC model for Eastern-language coverage |
| GigaAM | 2 | ~3.5 GB | Russian | Best-in-class Russian (CTC / RNNT) |
| Kaldi / Vosk / Zipformer | 5 | 93 – 478 MB | English + Russian + Arabic | Lightweight Vosk and Zipformer models for low-end hardware, incl. Quranic Arabic phonemes |
| T-One | 1 | 144 MB | Russian | Compact Russian streaming ASR |
A few standouts
| Model | Params | WER | RTFx | Languages | Default size |
|---|---|---|---|---|---|
| Whisper Tiny | 38M | 16.8 | 1350 | 99 | 149 MB |
| Whisper Large v3 Turbo | 796M | 7.0 | 783 | 99 | 1.1 GB |
| Moonshine Base | 62M | 8.6 | 2767 | EN | 711 MB |
| Parakeet TDT 0.6B v3 | 627M | 5.7 | 6098 | 25 | 2.5 GB |
| Canary 1B v2 | 978M | 6.4 | 1821 | 25 | 4.0 GB |
| Cohere Transcribe | 2.0B | 5.2 | 916 | 14 | 5.1 GB |
| Qwen3-ASR 0.6B | 600M | 5.6 | 439 | 30 | 2.0 GB |
| ARK-ASR 3B | 4.06B | 4.8 | 491 | 19 | 4.4 GB |
| Granite Speech 4.1 2B Plus | 2.0B | 4.9 | 547 | 5 | 4.1 GB int8 |
| SenseVoice Small | 234M | 10.6 | 850 | 5 | 159 MB |
How to pick
There is no single best model — it's a trade-off between accuracy, speed, size, and language. Start from what matters most to you.
Maximum accuracy
ARK-ASR 3B (WER 4.8) now leads, with Granite Speech 4.1 2B Plus (4.9) and Cohere Transcribe (5.2) close behind, or Whisper Large v3 for the widest language reach.
Lowest latency
Moonshine (up to ~3700 RTFx) is built for real-time English on CPU. For accurate English fast, Parakeet clears 5000 RTFx.
Smallest footprint
Whisper Tiny (149 MB), SenseVoice Small (159 MB), or Moonshine Tiny ship the smallest defaults; q4 / int8 quantizations shrink them further.
Specific language
Whisper covers 99 languages; Qwen3-ASR 30; Cohere 14; Canary / Parakeet v3 25 European languages; GigaAM / Vosk / T-One are Russian; Moonshine has per-language fine-tunes.
Good default
If you're unsure, start with Whisper Large v3 Turbo (multilingual, WER 7.8, ~1.1 GB) or Parakeet TDT 0.6B v3 for fast, accurate English and European dictation. Small machines can begin with Whisper Tiny or Moonshine Base and grow from there.
Real-time preview needs a small model
The live preview only runs models under ~700M parameters. Pair a large main model with a small realtime model (Whisper Tiny, Moonshine, or Canary 180M Flash) so words still stream as you speak.
Quantizations and downloads
Most models offer several quantizations — compressed weight formats (fp16, int8,
q4, q4f16, bnb4) that trade a little accuracy for big wins in size and speed. The
picker only shows the variants the upstream repo actually publishes, and you can download,
pause, resume, or delete each one independently. See
Compute & quantization for the full breakdown.
Everything stays local
Models download once into your user data folder and run entirely on-device. Audio is processed in memory and never uploaded — switching models never changes that. The only exception is the optional cloud STT providers, which you opt into explicitly with an API key.
Related
Whisper & Lite-Whisper
OpenAI Whisper, the EN-only variants, and the compressed Lite-Whisper builds.
NeMo Parakeet & Canary
NVIDIA's high-accuracy Conformer models for English, multilingual, and Russian.
Other models
Moonshine, Cohere, Granite, Qwen3-ASR, Audio8-ASR, SenseVoice, Dolphin, GigaAM, Vosk/Kaldi, and T-One.
Compute & quantization
What fp16, int8, q4, and q4f16 mean — and how to read the picker's badges.
Integrations
Add API keys to unlock optional cloud speech-to-text (ElevenLabs) and LLM cleanup (Ollama, OpenRouter). Keys are verified, stored encrypted on your device, and only ever sent to the provider you choose.
Whisper & Lite-Whisper
OpenAI Whisper — tiny through large-v3 — plus the compressed Lite-Whisper variants and the CrisperWhisper / Breeze fine-tunes, all on the ONNX engine.