Whisper & Lite-Whisper

Whisper is OpenAI's general-purpose speech recognition model, and the Whisper family is the most flexible engine WinSTT ships: 13 entries spanning a 38M-parameter realtime model up to a 1.55B-parameter accuracy ceiling, covering 99 languages with optional one-pass translation to English.

Models in the family

Languages, with auto-detect

38M–1.55B

Parameter span

7.4

Accuracy ceiling (large-v3 WER)

WinSTT runs every Whisper model through the same onnx-asr ONNX Runtime engine as all other families — not CTranslate2/faster-whisper. There is no backend to choose; pick a model in the Model tab and it loads on the one engine.

The WinSTT model picker open over the main window, showing selectable Whisper models with their accuracy and speed bars. — Pick any Whisper model in the Model tab. The accuracy and speed bars map to the WER and RTFx figures below.

How to read the numbers

Every model row below carries four catalog figures. They are the same numbers the model picker shows as accuracy/speed bars.

WER and RTFx are anchored to the Hugging Face Open ASR Leaderboard (English short-form, H200); rows the leaderboard doesn't carry are calibrated estimates on the same scale. RTFx is engine-and-hardware dependent; treat it as a relative ranking, not an absolute throughput guarantee.
Field	Means	Direction
Params	Parameter count — drives RAM/VRAM and download size.	Lower = lighter
WER	Word error rate on a mixed multilingual benchmark, as a percentage.	Lower = more accurate
RTFx	Real-time factor — how many seconds of audio it transcribes per wall-clock second.	Higher = faster
Langs	Number of languages the model can transcribe.	Higher = broader

All Whisper-family models support the realtime live preview, so any of them can serve as the realtime model as well as the main model.

Multilingual models

These cover all 99 Whisper languages with automatic language detection. Leave Language on auto-detect, or pin one in the Model tab.

large-v3-turbo has ~half the parameters of large-v3, runs ~1.7× faster, and trails it by only ~0.5 WER — it is the best all-round main model in the family.
Model	Params	WER	RTFx	Langs	Default size
tiny	38M	16.8	1350	99	149 MB
base	73M	12.5	1250	99	262 MB
small	242M	8.6	1050	99	772 MB
medium	769M	7.0	710	99	2.29 GB
large-v3-turbo	796M	7.0	783	99	1.13 GB
large-v3	1.55B	6.6	462	99	4.54 GB

English-only models

The .en variants drop the language-detection step. They are single-language (English) and a little more accurate on English audio than their multilingual twin at the same size — note each .en row's WER is lower than the multilingual one above it.

The Language selector is hidden for .en models — they are English-only by design and never auto-detect.
Model	Params	WER	RTFx	Default size
tiny.en	38M	14.6	1350	149 MB
base.en	74M	10.2	1250	262 MB
small.en	244M	7.4	1050	772 MB
medium.en	769M	6.3	710	4.33 GB

Lite-Whisper

Lite-Whisper is Efficient-Speech's structurally compressed large-v3-turbo. It keeps the full 99-language multilingual coverage but with a lower-rank encoder, so it downloads smaller and needs less memory than turbo. Note the compression pays off in footprint, not throughput — on the leaderboard hardware all three variants run at roughly the same speed as each other and slower than turbo itself. Three points on the size/accuracy curve:

All three are 99-language multilingual and support realtime. Only -acc is independently measured on the leaderboard; the other two rows are calibrated from upstream's ESB results onto the same scale.
Model	Params	WER	RTFx	Trade-off
lite-whisper-large-v3-turbo-acc	581M	6.5	327	Least compressed — most accuracy retained
lite-whisper-large-v3-turbo	534M	8.1	335	Balanced compression
lite-whisper-large-v3-turbo-fast	474M	12.9	340	Most aggressive compression

The -acc suffix means accuracy-preserving, not accelerated

lite-whisper-large-v3-turbo-acc (shown as Accurate in the picker) is the least compressed of the three — it keeps the most parameters and essentially matches turbo's accuracy in ~60% of the size. The plain (no-suffix) variant is the balanced middle, and -fast is the smallest, at a real accuracy cost. Read by the numbers above, not the suffix.

Fine-tunes

Two specialist Whisper-large fine-tunes for narrower jobs. Both are 1.5B-class, multilingual-capable, and ship in default precision only.

CrisperWhisper 2.0 is a Whisper-large fine-tune (Nyra Labs, non-commercial research license) that tops the Nyra Verbatim Speech Benchmark at 87.8 disfluency F1 across ten languages; Breeze ASR 25 is a large-v2 fine-tune that retains the full multilingual vocabulary.
Model	Params	WER	RTFx	Langs	Specialty
crisper-whisper	1.54B	5.8	33	99	CrisperWhisper 2.0 — verbatim transcription with ~30 ms word timing. Keeps fillers ([um]/[uh]), stutters, and vocal events ([laughter], [cough]) instead of cleaning them up. Multilingual.
crisper-whisper-turbo	809M	6.5	100	99	CrisperWhisper 2.0 Turbo — the same verbatim + word-timing behavior on the large-v3-turbo architecture (4 decoder layers): much faster decodes at a small accuracy cost.
breeze-asr-25	1.55B	8.0	460	99	MediaTek Breeze — a large-v2 fine-tune tuned for Mandarin/English code-switching.

Verbatim vs. clean

Most ASR models silently delete disfluencies. CrisperWhisper 2.0 does the opposite — it transcribes exactly what was said, fillers and all, as bracketed event markers ([um], [laughter]). Use it for interview/legal transcripts where every "um" matters; for dictation, enable Remove Filler Words (Quality → Formatting) to strip the markers while keeping its accuracy.

Translate to English

Multilingual Whisper models support OpenAI's native task=translate. Turn on Translate to English in the Model tab and the model transcribes and translates to English in a single decode pass — no extra latency and no LLM round-trip.

Translate is a Whisper-native task; non-Whisper families translate via the LLM cleanup pass instead.
Model class	Translate to English	Note
Multilingual (tiny … large-v3, lite-whisper-*, breeze-asr-25)	Supported	Single-pass; toggle in the Model tab.
English-only (.en)	N/A	Source is already English.
crisper-whisper (en/de)	Supported	German source → English output.

Precision (quantizations)

Whisper precision is set per-model via the picker's quantization badge (see Compute Types). The empty default ("") is fp32 — maximum accuracy. The picker only offers what the upstream repo actually publishes.

fp16 roughly halves the download and is the fastest path on DirectML GPUs; q4 / bnb4 cut size to ~13% for the smallest footprint. The server falls back to default precision if a chosen variant is missing.
Models	Available quantizations
tiny, base, small, medium, large-v3-turbo, and all .en variants	default · fp16 · q4 · bnb4
lite-whisper-large-v3-turbo / -acc / -fast	default · fp16
large-v3, crisper-whisper, breeze-asr-25	default only

Requires a restart

Switching between Whisper models hot-swaps live without a restart. Changing to a different model family (e.g. Whisper → NeMo) reconnects the engine — the overlay shows a brief "swapping" state.

When to use which

Realtime preview

tiny or base (or .en for English) — 1250–1350 RTFx keeps the live preview snappy.

Best all-round main model

large-v3-turbo — 7.0 WER at 783 RTFx and a 1.13 GB download. The default recommendation.

Maximum accuracy

large-v3 — 6.6 WER, the family's accuracy ceiling, at the cost of a 4.54 GB download and roughly half turbo's speed.

Turbo quality, smaller footprint

lite-whisper-large-v3-turbo-acc — 6.5 WER, 581M params, full multilingual coverage.

English-only boost

The .en variants shave 1–2 WER off English audio versus their multilingual twins.

Verbatim transcripts

crisper-whisper — keeps every filler and false start for interview and legal work.

Recommended pairing

Run large-v3-turbo as the main model for the final, accurate paste, and pair it with tiny or base as the realtime model for a responsive live preview. If your main model is already small enough to run live, WinSTT locks one model into both slots automatically.