bartelds/whisper-dro-set2-dro

Whisper CTC-DRO ASR model - set 2

This repository contains an automatic speech recognition (ASR) model fine-tuned from openai/whisper-large-v3 using the principles of CTC-DRO applied to Whisper's seq2seq architecture. The model was trained on balanced training data from set 2 (eng, fas, hrv, ita, slk, yue).

DRO hyperparameters: eta=5e-3, alpha=0.1, aggregation: mean

Intended Use

This model is intended for multilingual ASR. Users can run inference using the HuggingFace Transformers library:

import torch
import librosa
from transformers import WhisperForConditionalGeneration, WhisperProcessor

model = WhisperForConditionalGeneration.from_pretrained("bartelds/whisper-dro-set2-dro")
processor = WhisperProcessor.from_pretrained("bartelds/whisper-dro-set2-dro")
model.eval()

audio, sr = librosa.load("input.wav", sr=16000)
inputs = processor.feature_extractor(audio, sampling_rate=16000, return_tensors="pt")

with torch.no_grad():
    generated = model.generate(input_features=inputs.input_features)

text = processor.tokenizer.batch_decode(generated, skip_special_tokens=True)[0]
print("Recognized text:", text)

How to Use

1. Install dependencies: pip install transformers torch librosa
2. Load the model and processor using from_pretrained() as shown above.
3. The model supports multilingual transcription -- see the training repository for evaluation details.

Training

• Base model: openai/whisper-large-v3
• Training code: whisper-dro
• Paper: CTC-DRO