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.

Moonshine

Moonshine is a raw-audio encoder built for latency. The catalog ships Moonshine v2 ("Ergodic Streaming Encoder"): its sliding-window encoder runs incrementally — each 40 ms hop of audio is encoded once as it arrives, so live text appears while you speak and the final transcript is ready almost as soon as you stop. Tiny, Small and Medium cover English; the language tunes are single-language only, with no auto-detect. Three first-generation (v1) tunes stay for languages v2 does not cover yet (Korean, Ukrainian, French).

v2 models ship fp32 ("", default) and int8 graphs; int8 is 20–30% faster on CPU at near-identical accuracy. The v1 tunes keep their wider quantization range (fp16, q4, bnb4, int8, uint8, q4f16). Moonshine runs on the CPU.

The v1 English, Arabic, Vietnamese, Chinese and Japanese models were replaced; a saved selection moves to its v2 counterpart automatically (Base → v2 Small, Base JA → v2 Small JA, Base ZH → v2 Tiny ZH). Korean and Ukrainian v1 tunes are under the Moonshine Community License (non-commercial; free commercial use under US$1M annual revenue); everything else is MIT.

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.

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.

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.

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.

Audio8-ASR and ARK-ASR

Audio8 ships two arkasr models (plus Edge0's streaming Infinite model, below). 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 Infinite (Edge0, Apache-2.0) is a different architecture under the Audio8 name: a causal, Voxtral-style audio tower feeding a Qwen2.5-3B decoder that emits one token every 80 ms at a fixed 480 ms delay. It is a native streaming model — live dictation sees text grow word by word instead of re-decoding a window — and both attention caches roll, so memory stays constant however long the recording runs (the 24-second chunking above does not apply). The decoder also predicts end of turn, which WinSTT uses to finalize a segment. It covers Chinese and English only and cannot auto-detect, so set the spoken language explicitly — speech in the other language gets translated rather than transcribed. At ~4.1B parameters it runs on the CPU and needs a fast multi-core machine to keep up in real time.

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.

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.

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).

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.

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).

Picking by language

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

Related