Other Families

Beyond Whisper and NeMo, WinSTT ships ten more model families on the same ONNX engine — covering ultra-low-latency English dictation, leaderboard-topping multilingual ASR, LLM-decoder ASR, East-Asian and Eastern-language coverage, and three Russian-specialist engines. Pick by language coverage first, then by the accuracy/speed trade-off.

Every model on this page runs through the single onnx-asr ONNX engine, supports the live realtime preview, and is hot-swappable from the Model tab — no restart, no backend to choose. Precision is set per-model via onnxQuantization; see Compute Types.

model families on this page

4.9

lowest WER (Granite Speech 4.1 2B Plus)

3733

fastest RTFx (Moonshine Tiny)

Qwen3-ASR languages

Moonshine

Useful Sensors Moonshine is a raw-audio encoder built for latency, not raw accuracy. The tiny model is ~2.8× faster than Whisper tiny on the leaderboard benchmark, which makes it a strong realtime-preview and live-dictation pick. The base English models and the language fine-tunes round out the family — each fine-tune is single-language only, with no auto-detect.

RTFx = real-time factor (higher is faster); the fine-tunes share their base model's architecture and speed. WER is whichever benchmark the upstream repo reports — read it within a family, not across.
Model	Params	WER	RTFx	Language
Moonshine Tiny	27M	11.4	3733	English
Moonshine Base	62M	8.6	2767	English
Moonshine Tiny (KO)	27M	15.0	3733	Korean
Moonshine Tiny (AR)	27M	15.0	3733	Arabic
Moonshine Tiny (VI)	27M	15.0	3733	Vietnamese
Moonshine Tiny (UK)	28M	—	3733	Ukrainian
Moonshine Tiny (FR)	28M	—	3733	French
Moonshine Base (ZH)	62M	10.5	2767	Chinese
Moonshine Base (JA)	62M	10.5	2767	Japanese
Moonshine Base (KO)	62M	10.5	2767	Korean

Moonshine exposes the widest quantization range of any family: "" (default), fp16, q4, bnb4, int8, uint8, and q4f16. The int8 / uint8 graphs are the smallest downloads (the English tiny model drops from 327 MB to 67 MB) and are the best fit for CPU-only machines.

The fastest realtime model

At 27M parameters and RTFx ~3700, Moonshine Tiny is the lightest model in the catalog and an ideal realtime preview model. Pair it with a larger main model when you want the live preview to feel instant without taxing the GPU.

Cohere

Cohere Transcribe is a 2B-parameter Conformer encoder with a lightweight Transformer decoder. It sits near the top of the Hugging Face Open ASR Leaderboard at 5.2 average WER — second in the catalog only to Granite Speech 4.1 2B Plus. It covers 14 languages with automatic language detection.

Languages: Arabic, German, Greek, English, Spanish, French, Italian, Japanese, Korean, Dutch, Polish, Portuguese, Vietnamese, Chinese — with auto-detect.
Model	Params	WER	RTFx	Languages
Cohere Transcribe	2B	5.2	916	14

Quantizations: "" (default), fp16, q4, and q4f16. The default fp32 graph is the largest single download in the catalog (~5 GB); on a GPU, the fp16 (~2.4 GB) or q4f16 (~1.5 GB) variants give you most of the accuracy for a fraction of the size and VRAM.

Plan for the download and VRAM

Cohere is a 2B model. Even the smallest q4f16 download is ~1.5 GB, and the default precision needs roughly 5 GB of free VRAM to create the GPU session. If the session can't fit, ONNX Runtime falls back to CPU — accurate, but slow at this size.

Granite Speech

IBM Granite Speech 4.1 brings two 2B-parameter models to the catalog: an autoregressive Plus model and a non-autoregressive NAR model. Both cover English, French, German, Spanish, and Portuguese, and both are useful when punctuation and written-form output matter as much as raw word recognition.

The Plus model posts the lowest WER in the catalog; the single-pass NAR model gives up almost nothing in accuracy while decoding ~4× faster. The older Granite Speech 4.1 2B id is migrated to the newer 2B Plus export.
Model	Params	WER	RTFx	Languages	Quant	Typical size
Granite Speech 4.1 2B Plus	2B	4.9	547	en · fr · de · es · pt	—, int8, fp16w	4.1 GB int8 / 16.4 GB default
Granite Speech 4.1 2B NAR	2B	5.0	2079	en · fr · de · es · pt	—, int8, fp16w	2.5 GB int8 / 9.8 GB default

Large-model class

Granite is a large speech model family. Start with int8 unless you are explicitly testing the full export; the default graphs are much larger than typical dictation models.

Qwen3-ASR

Qwen3-ASR is the new Alibaba Qwen family in this release. It pairs a Whisper-style audio encoder with a Qwen3 decoder and runs through three ONNX graphs (encoder, decoder_init, decoder_step) plus a shared fp16 embedding table. The catalog ships the int4 exports because the full precision graphs are too large for normal desktop use.

Languages include Chinese, English, Cantonese, Arabic, German, French, Spanish, Portuguese, Indonesian, Italian, Korean, Russian, Thai, Vietnamese, Japanese, Turkish, Hindi, Malay, Dutch, Swedish, Danish, Finnish, Polish, Czech, Filipino, Persian, Greek, Hungarian, Macedonian, and Romanian.
Model	Params	Languages	Quant	Download
Qwen3-ASR 0.6B	600M	30 + auto-detect	int4 only	2.0 GB
Qwen3-ASR 1.7B	1.7B	30 + auto-detect	int4 only	4.1 GB

CPU-first for correctness

Qwen3-ASR is currently pinned away from unverified non-CUDA GPU paths. That keeps the new int4 LLM-decoder graph on the safest execution path while the engine matures.

Audio8-ASR and ARK-ASR

Audio8 ships two arkasr models. Audio8-ASR 0.1B is the smallest LLM-decoder model in the catalog: the same Qwen3-ASR audio tower, but feeding an 8-layer Qwen-style decoder of only 0.1B parameters (0.32B end-to-end). That makes it the one LLM-decoder option that fits comfortably on a CPU-only machine — the int4 download is under 700 MB, where Qwen3-ASR starts at 2.0 GB. ARK-ASR 0.6B trades that footprint for accuracy: a Whisper-large encoder over a 0.6B decoder (~1.3B end-to-end), 19 languages, and 5.53 WER on the Open ASR Leaderboard. ARK-ASR 3B pushes the same design to a 3B decoder and posts 4.76 WER — the most accurate model in the catalog — at the cost of a 4.4 GB download and noticeably slower CPU decoding.

Audio8-ASR covers English, Chinese, Cantonese, French, Japanese, German, and Korean; its int4 tier pairs the int8 audio tower with the int4 decoder (the bundle ships no int4 tower). ARK-ASR 0.6B is the larger sibling — a Whisper-large encoder over a 0.6B decoder, 19 languages, WER 5.53 — and is Apache-2.0, so the non-commercial note below applies to the 0.1B only.
Model	Params	Languages	Quant	Download
Audio8-ASR 0.1B	0.32B	7 + auto-detect	fp32 · int8 · int4	0.67 – 2.0 GB
ARK-ASR 0.6B	1.3B	19 + auto-detect	int8 only	1.8 GB
ARK-ASR 3B	4.06B	19 + auto-detect	int8 · int4	2.9 – 4.4 GB

On ARK-ASR 3B, int4 is smaller — not faster

Unlike the 0.1B (where int4 is the fastest tier), the 3B's int4 weights are ~50% slower than its int8 ones on CPU: 4-bit weights are dequantized on the fly, and at this size that costs more than the narrower weights save. Pick int8 for speed and int4 only when you need the 1.5 GB back.

Non-commercial licence

Audio8-ASR is released under CC BY-NC 4.0 — commercial use is not permitted under that licence. Check the model card before using it for anything beyond personal or research work. Cohere Transcribe carries the same restriction; every other family on this page is permissively licensed.

Decodes in 24-second chunks

The decoder's key/value cache is a fixed 512 positions shared between the audio prompt and the transcript, so WinSTT segments longer recordings on silence at 24 s. This is automatic — long recordings transcribe in full — but it means Audio8 re-encodes per chunk rather than holding a whole take in one pass.

SenseVoice

FunAudioLLM SenseVoice Small is a 234M Conformer encoder focused on East-Asian languages: Chinese, English, Japanese, Korean, and Cantonese, with automatic language detection. It ships as a single-pass int8 graph — there is no fp32 variant — so the download is a compact 159 MB and the only available precision is int8.

yue = Cantonese. SenseVoice ships int8-only; the picker offers no other quantization.
Model	Params	WER	RTFx	Languages
SenseVoice Small	234M	10.6	850	zh · en · ja · ko · yue

Dolphin

DataoceanAI Dolphin Base CTC is a compact int8 CTC model for Eastern-language coverage: Arabic, Chinese, Japanese, Korean, Vietnamese, Hindi, Thai, Indonesian, Malay, Tamil, Urdu, Persian, Bengali, Nepali, and Burmese. It is a good alternative when Whisper's broad multilingual model is more than you need.

Dolphin is single-pass CTC and ships only as an int8 graph.
Model	Params	Quant	Download	Languages
Dolphin Base CTC	140M	int8 only	104 MB	15

Russian-specialist families

Three families target Russian exclusively. They are the right pick when you dictate primarily in Russian and want accuracy that beats Whisper at a fraction of the size. (NVIDIA's Russian FastConformer models live in the NeMo family.)

GigaAM

GigaAM v3 is SaluteDevices' 243M Russian engine, shipped in two end-to-end variants: a CTC graph (faster decoding) and an RNN-T graph (higher accuracy).

Russian only. Both variants offer the default precision plus int8. WER is the upstream multi-benchmark average — compare it within this family, not against the other Russian engines' single-benchmark figures.
Model	Params	WER	RTFx	Architecture
GigaAM v3 E2E CTC	243M	9.2	5000	CTC — faster decoding
GigaAM v3 E2E RNNT	243M	8.4	4500	RNN-T — higher accuracy

GigaAM supports "" (default) and int8. The default graph is a ~3.5 GB download; int8 drops it to ~900 MB.

Vosk / Kaldi

Vosk provides lightweight offline recognition built on the Kaldi toolkit. Both Russian models are small, load fast, and run comfortably on CPU — the small variant is the smallest Russian option in the catalog.

Russian only. Both offer the default precision plus int8.
Model	Params	WER	RTFx	Notes
Vosk Russian	65M	6.1	350	Full model — higher accuracy
Vosk Russian (Small)	23M	9.8	800	Lightweight — faster, smaller download

Vosk/Kaldi models support "" (default) and int8. The full model is a 264 MB download (72 MB at int8); the small model is 93 MB (27 MB at int8).

T-One

T-Tech's T-One is a 72M Russian engine that hits accuracy on par with the NeMo Russian FastConformer models at a much smaller footprint (144 MB).

Russian only. Ships in default precision — the picker offers no int8 variant.
Model	Params	WER	RTFx	Notes
T-One	72M	5.3	500	Default precision only

Picking by language

Use this to narrow the list before weighing accuracy and speed:

For broad multilingual coverage with auto-detect, also compare the Whisper and NeMo families.
You need…	Best fit here
Fastest English realtime / dictation	Moonshine Tiny or Base
Top accuracy, 14 languages	Cohere Transcribe
LLM-decoder multilingual ASR	Qwen3-ASR 0.6B or 1.7B
LLM-decoder ASR on a CPU-only machine	Audio8-ASR 0.1B
Written-form punctuation in five languages	Granite Speech 4.1
Chinese / Japanese / Korean / Cantonese	SenseVoice Small (or a Moonshine fine-tune)
Eastern-language CTC coverage	Dolphin Base CTC
Arabic or Vietnamese	Moonshine Tiny (AR) / (VI)
Best Russian accuracy	GigaAM v3 RNN-T
Smallest Russian, CPU-only	Vosk Russian (Small) or T-One

Single-language models hide the language picker

The Moonshine fine-tunes and every Russian model are single-language — the Language control in the Model tab disappears for them, since there is nothing to detect or pin. Cohere, Qwen3-ASR, Audio8-ASR, and SenseVoice support auto-detect, so they keep it.

Hot-swapping vs. family switches

Switching between models within this group hot-swaps live without a restart. Changing to a different model family (e.g. Moonshine → Qwen3-ASR) reconnects the engine — the overlay shows a brief "swapping" state.