whatmodel.sh / stt / ESPnet Branchformer 18-layer (70 epochs, no LM)

ESPnet Branchformer 18-layer (70 epochs, no LM)

+ contribute
Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

ESPnet Branchformer 18-layer (70 epochs, no LM) is an open-weights speech-to-text model from ESPnet, released June 2022. Its published size is 0.11688B parameters. Licence: cc-by-4.0.

Catalogue identity notes

Exact LibriSpeech 960h recipe run, README lines 319–399. Published checkpoint linked by recipe authors; a compatible local command has not yet been verified. Decoder and checkpoint averaging are recorded per result. Architecture names do not denote a generic pretrained model.

Model information sources

· Parameter count ↗

README lines 319-320 (identity note cites 319-399); model link matches model_repo. ASR model only; LM not included.

44.0±3.9#127 of 162 speech-to-text modelsfrom 2 benchmark results across 1 source* thinly tested, treat the rank as provisional

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 ESPnet Branchformer 18-layer (70 epochs, no LM), out of 176 with a score, then more from ESPnet.

#modelscorewhereRAMreleased
138Distil-Whisper distil-medium.en Hugging Face (Distil-Whisper)44±3local~1.4GB2023-10
139Whisper small.en OpenAI44±3local~1.1GB2022-09
140ESPnet Branchformer 18-layer (70 epochs, no LM) ESPnet44±4local~0.7GB2022-06
141Nemotron 3.5 ASR Streaming 0.6B NVIDIA44±3local~0.7GB2026-05
142Whisper medium OpenAI43±2local~2.3GB2022-09
more from ESPnet
112ESPnet HuBERT Large + Conformer + Transformer LM ESPnet50±4local~8GB2022-01
113ESPnet WavLM Large + Conformer + Transformer LM ESPnet50±4local~8GB2022-01
129ESPnet E-Branchformer (no LM) ESPnet46±4local~0.8GB2023-01

see all speech-to-text models · find the ones that fit your machine

$ what it's good at

other tests2 not in a task · 94% of its score
testresultvs the others in its rundifficultycountswhy
LibriSpeech test-other WER lab5.3% ✎█▃▃·▃▅·▅···▅ beats 5/11avg 3.7%▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 73%typical
LibriSpeech test-clean WER lab2.3% ✎█··▃▃▅▅▃···▅ beats 6/11avg 1.9%▇▇▇▇▇▇▇▇ 21%blunt

$ where to run it

Other ways to run it

  1. Run it as published · needs ~0.68 GB · from github.com/espnet (official)
    uv run --python 3.12 --with espnet==202610.post2 --with espnet_model_zoo==0.1.11 --with librosa python -c "import librosa; from espnet2.bin.asr_inference import Speech2Text; asr=Speech2Text.from_pretrained('pyf98/librispeech_branchformer_e18_linear3072@f37570da7c3e4c6432e1ce29963f39c0c8127c83', beam_size=60, ctc_weight=0.3, lm_weight=0); speech,_=librosa.load('audio.wav', sr=16000, mono=True); print(asr(speech)[0][0])"

Weights and documentation: huggingface.co/pyf98 ↗

$ how the score was worked out · ESPnet LibriSpeech ASR recipe… 94%

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

sourceresultsshare of the scorelicence
ESPnet LibriSpeech ASR recipe results (pinned September 2026) · lab-reported2▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 94%Apache-2.0
prior · pull toward the middle▇▇ 6%
weighted average= 44.0

$ improve this page

Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

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