WinSTTNVIDIA NeMo
NVIDIA's Parakeet and Canary models — Conformer-based recognizers that run at 1800–6100× real-time on the same ONNX engine as everything else.
NVIDIA's NeMo family pairs FastConformer encoders with three different decoders (CTC, RNN-T, TDT) to hit some of the highest accuracy-per-millisecond numbers in the catalog — Parakeet TDT lands at 5.7 WER while decoding at ~6100× real-time. WinSTT runs every NeMo model on the same onnx-asr ONNX Runtime engine as Whisper and the rest; there is no separate backend to flip.
The line-up
Two product lines share the FastConformer encoder:
- Parakeet — single-pass recognizers tuned for raw speed. English-only for CTC/RNN-T, 25-language for the newer TDT.
- Canary — encoder-decoder models with a small attention decoder. They add automatic language detection and Whisper-style
translatesupport, trading some throughput for it.
A pair of FastConformer RU models round out Russian coverage.
| Model | Params | WER | RTFx | Languages | Quant |
|---|---|---|---|---|---|
| Parakeet CTC 0.6B | 600M | 6.7 | 5884 | English | —, int8 |
| Parakeet RNNT 0.6B | 600M | 6.7 | 5407 | English | —, int8 |
| Parakeet TDT 0.6B v3 | 627M | 5.7 | 6098 | 25 (multilingual) | —, int8 |
| Canary 180M Flash | 194M | 6.3 | 2484 | en · de · es · fr | —, int8 |
| Canary 1B Flash | 883M | 5.8 | 2126 | en · de · es · fr | —, int8 |
| Canary 1B v2 | 978M | 6.4 | 1821 | 25 (multilingual) | —, int8 |
| FastConformer RU CTC | 109M | 5.3 | 5700 | Russian | —, int8 |
| FastConformer RU RNNT | 114M | 5.3 | 5400 | Russian | —, int8 |
Every NeMo model supports the live preview (realtime transcription), so any of them can serve as both your main and realtime model.
CTC vs RNN-T vs TDT
All three decoders sit on top of the same FastConformer encoder. They differ only in how they turn encoder frames into text — which is what moves the speed/accuracy trade-off.
| Decoder | How it decodes | Trade-off |
|---|---|---|
| CTC | Each audio frame is classified independently in a single pass, then collapsed. | Fastest, lightest. Slightly higher WER because it can't condition on what it just emitted. |
| RNN-T | Autoregressive — each token is predicted from the audio and the tokens emitted so far. | More accurate than CTC; the token-by-token loop is heavier, so RTFx is lower. |
| TDT | An RNN-T variant that predicts a token and its duration together, so it can skip ahead instead of stepping one frame at a time. | RNN-T accuracy at close-to-CTC speed — NVIDIA's newest decoder, and why TDT v3 beats both 0.6B Parakeets. |
The Russian pair publishes the same 5.3 WER for both decoders; the CTC graph (5700 RTFx) decodes a touch faster than the RNN-T one (5400 RTFx).
English-only vs multilingual
Language scope is fixed per model — it is not a setting you can widen.
- English-only:
Parakeet CTC 0.6BandParakeet RNNT 0.6B. The fastest path to high-accuracy English dictation; no language-detection step to pay for. - Four-language (en / de / es / fr): both Canary Flash models.
180M Flashis the smallest model in the family (2484 RTFx);1B Flashkeeps Flash-class speed (2126 RTFx) at noticeably lower WER (5.8). - 25-language:
Parakeet TDT 0.6B v3andCanary 1B v2both cover the same 25 European languages (Bulgarian, Czech, Danish, German, Greek, English, Spanish, Estonian, Finnish, French, Croatian, Hungarian, Italian, Lithuanian, Latvian, Maltese, Dutch, Polish, Portuguese, Romanian, Russian, Slovak, Slovenian, Swedish, Ukrainian). Both auto-detect the spoken language. - Russian-only: the two
FastConformer RUmodels.
TDT v3 or Canary 1B v2 for multilingual?
They cover the same 25 languages. TDT v3 is the right default — 5.7 WER at 6098 RTFx in 2.5 GB, more accurate and ~3× faster on the leaderboard average. Pick Canary 1B v2 (6.4 WER, 1821 RTFx, 4 GB) when you need its real translate task.
Precision
Every NeMo model ships two quantizations, selected by the Quantization control on the Model tab:
| Value | Precision | Best for |
|---|---|---|
| — (default) | Full precision | Maximum accuracy; GPU |
| int8 | 8-bit integer | CPU — ~3.5–4× smaller download and lower RAM |
For example, Parakeet TDT 0.6B v3 is 2.5 GB at default precision and 670 MB as int8; Canary 1B v2 drops from 4.0 GB to 1.0 GB. See Compute Types for what the quantization labels mean.
No backend to set
Every model — NeMo included — runs on the single ONNX engine. There's no "NeMo backend" toggle; just pick a NeMo model in Model and it hot-swaps in. Switching to or from a different model family briefly reconnects the engine.
When to use which
Fastest English dictation
Parakeet CTC 0.6B — 6.7 WER at 5884 RTFx, single-pass, English-only. No language-detection step to pay for.
Best all-round main model
Parakeet TDT 0.6B v3 — 5.7 WER at 6098 RTFx in 2.5 GB, covering 25 languages. The family's accuracy ceiling and its throughput champion at once — the default multilingual pick.
Built-in translation
Canary 1B v2 — the only NeMo model with a real translate task. 6.4 WER, 4 GB, 1821 RTFx.
Smallest download
Canary 180M Flash — 6.3 WER at 2484 RTFx from a sub-200M model, across en / de / es / fr.
Four-language balance
Canary 1B Flash — 5.8 WER at 2126 RTFx across en / de / es / fr.
Russian-only
FastConformer RU RNNT — 5.3 WER for Russian dictation; the CTC sibling matches it and decodes slightly faster.
Related
Browse all families
Whisper, NeMo, Moonshine, Cohere, GigaAM, Vosk, and more across the full catalog.
Whisper Models
OpenAI Whisper and the compressed Lite-Whisper variants — multilingual and English-only.
Compute & quantization
What int8, fp16, and q4 mean, and how to read the quantization labels.
Pick a model
Choose your main and realtime models, language, device, and precision.
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.
Other Families
Moonshine, Cohere, Granite, Qwen3-ASR, Audio8-ASR, SenseVoice, Dolphin, GigaAM, Vosk/Kaldi, and T-One — the low-latency, multilingual, and Russian-focused models beyond Whisper and NeMo.