whatmodel.sh / stt / LiteWhisper large-v3
LiteWhisper large-v3
↗ share
+ contribute
LiteWhisper large-v3 is an open-weights speech-to-text model from Efficient-Speech (LiteASR), released February 2025. Its published size is 1.35B parameters. Licence: apache-2.0.
Catalogue identity notes
Exact checkpoint efficient-speech/lite-whisper-large-v3 (low-rank compressed Whisper encoder; custom code, trust_remote_code). Params: HF safetensors total. released: HF repo creation month. Command follows the card snippet (processor from openai/whisper-large-v3).
Score within the Open ASR English short-form WER (March 2026) comparison set only.
$ how it compares
The speech-to-text models ranked either side of LiteWhisper large-v3, out of 176 with a score, then more from Efficient-Speech (LiteASR).
| # | model | score | where | ||
|---|---|---|---|---|---|
| 77 |  Distil-Whisper distil-large-v3 | 56±2 | local | ||
| 78 |  Omnilingual ASR LLM 3B v2 | 56±3 | local | ||
| 79 | LiteWhisper large-v3 | 55±3 | local | ||
| 80 |  Fun-ASR-Nano-2512 | 55±3 | local | ||
| 81 |  Scribe v2 Realtime | 55±3 | api $0.0065/audio min | ||
| more from Efficient-Speech (LiteASR) | |||||
| 82 | LiteWhisper large-v3-acc | 55±3 | local | ||
| 95 | LiteWhisper large-v3-turbo-acc | 53±4 | local | ||
| 101 | LiteWhisper large-v3-fast | 52±3 | local | ||
see all speech-to-text models · find the ones that fit your machine
$ what it's good at
other tests4 not in a task · 96% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Open ASR English short-form WER (March 2026) | 7.4% ✎ | ▁▁▁▃··▁▁▂██▆ beats 50/86 | 33% |
| Multilingual LibriSpeech French test WER lab | 6.8% ✎ | █·······██·█ beats 3/3 | 25% |
| LiteASR paper · ESB 8-set average WER lab | 10.2% ✎ | ▂·······▂··█ beats 3/6 | 22% |
| Multilingual LibriSpeech German test WER lab | 7.7% ✎ | █····█···█·█ beats 3/3 | 17% |
$ where to run it
Other ways to run it
- Run it as published · needs ~3.74 GB, NVIDIA only · from huggingface.co/efficient-speech (official)
uv run --with transformers --with torch --with librosa python -c "import librosa, torch; from transformers import AutoProcessor, AutoModel; m=AutoModel.from_pretrained('efficient-speech/lite-whisper-large-v3', trust_remote_code=True).to(torch.float16).to('cuda'); p=AutoProcessor.from_pretrained('openai/whisper-large-v3'); f=p(librosa.load('audio.wav', sr=16000)[0], sampling_rate=16000, return_tensors='pt').input_features.to(torch.float16).to('cuda'); print(p.batch_decode(m.generate(f), skip_special_tokens=True)[0])"
Official weights and documentation: huggingface.co/efficient-speech ↗
$ how the score was worked out · LiteASR (arXiv:2502.20583v2)… 64%, Open ASR Leaderboard 33%
Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works
| source | results | share of the score |
|---|---|---|
 LiteASR (arXiv:2502.20583v2) · ESB WER of Whisper and LiteASR-compressed encoders · lab-reported | 3 | 64% |
 Open ASR Leaderboard | 1 | 33% |
| prior · pull toward the middle | 4% | |
| weighted average | = 55.4 |
$ improve this page
Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.
data updated · sources and licences · json