This is an openai/whisper-medium.en model whose encoder was domain-adapted with Gumbel-BEARD — a self-supervised continued-pretraining recipe — and then fine-tuned for automatic speech recognition (ASR) on the MyST children's speech corpus.

On the MyST test set it achieves 8.2% WER.

Usage

python

from transformers import pipeline

asr = pipeline(
    "automatic-speech-recognition",
    model="Zilai2999/whisper-medium.en-gumbel-beard",
)
print(asr("audio.wav")["text"])

Or with the model/processor directly:

python

import torch, librosa
from transformers import WhisperForConditionalGeneration, WhisperProcessor

model_id = "Zilai2999/whisper-medium.en-gumbel-beard"
processor = WhisperProcessor.from_pretrained(model_id)
model = WhisperForConditionalGeneration.from_pretrained(model_id)

audio, _ = librosa.load("audio.wav", sr=16000)
inputs = processor(audio, sampling_rate=16000, return_tensors="pt")
ids = model.generate(inputs.input_features)
print(processor.batch_decode(ids, skip_special_tokens=True)[0])

Audio is expected at 16 kHz.

Training

Self-supervised Gumbel-BEARD adaptation of the Whisper encoder (BEST-RQ + self-distillation with automatic Gumbel-softmax layer selection).
Supervised ASR fine-tuning on the MyST children's speech corpus.

See the code repository for the full recipe and hyper-parameters.

Results

Test set	WER
MyST	8.2%

Citation

bibtex

@inproceedings{gumbelbeard2026,
  title     = {Gumbel-BEARD: Automatic Layer Selection for Self-Supervised
               Adaptation of Whisper in Low-Resource Domains},
  booktitle = {Proc. Interspeech},
  year      = {2026},
}