EBranchRegulaFormer
This is a 174M encoder-decoder Ebranchformer model trained with an intermediate regularization technique on 6,000 hours of open-source English data.
It achieves Word Error Rates (WERs) comparable to openai/whisper-medium.en
across multiple datasets with just 1/4 of the parameters.
Architecture details, training hyperparameters, and a description of the proposed technique will be added soon.
Disclaimer: The model currently hallucinates on segments containing silence only, as it was previously not trained on such data. The fix will be added soon.
The model can be used with the pipeline
class to transcribe audio files of arbitrary length.
from transformers import pipeline
model_id = "BUT-FIT/EBranchRegulaFormer-medium"
pipe = pipeline("automatic-speech-recognition", model=model_id, feature_extractor=model_id, trust_remote_code=True)
# In newer versions of transformers (>4.31.0), there is a bug in the pipeline inference type.
# The warning can be ignored.
pipe.type = "seq2seq"
# Standard greedy decoding
result = pipe("audio.wav")
# Beam search decoding with joint CTC-autoregressive scorer
generation_config = pipe.model.generation_config
generation_config.ctc_weight = 0.3
generation_config.num_beams = 5
generation_config.ctc_margin = 0
result = pipe("audio.wav")
- Downloads last month
- 7
Inference API (serverless) is not available, repository is disabled.
Datasets used to train BUT-FIT/EBranchRegulaFormer-medium
Evaluation results
- Test WER on LibriSpeech (clean)test set self-reported2.500
- Test WER on LibriSpeech (clean)test set self-reported5.600
- Test WER on tedlium-v3test set self-reported6.300
- Test WER on Vox Populitest set self-reported7.300
- Test WER on Mozilla Common Voice 13.0test set self-reported12.100