VSydorskyi's picture
Update README.md
8c51d1f
metadata
language:
  - uk
license: apache-2.0
datasets:
  - mozilla-foundation/common_voice_11_0
model-index:
  - name: ukrainian-data2vec-asr
    results:
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Common Voice 11.0
          type: mozilla-foundation/common_voice_11_0
          config: uk
          split: test
          args: uk
        metrics:
          - name: Wer
            type: wer
            value: 17.04228333878635
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: Common Voice 11.0
          type: mozilla-foundation/common_voice_11_0
          config: uk
          split: validation
          args: uk
        metrics:
          - name: Wer
            type: wer
            value: 17.6343500009732

Respeecher/ukrainian-data2vec-asr

This model is a fine-tuned version of Respeecher/ukrainian-data2vec on the Common Voice 11.0 dataset Ukrainian Train part. It achieves the following results:

  • eval_wer: 17.634350000973198
  • test_wer: 17.042283338786351

How to Get Started with the Model

from transformers import AutoProcessor, Data2VecAudioForCTC
import torch
from datasets import load_dataset, Audio

dataset = load_dataset("mozilla-foundation/common_voice_11_0", "uk", split="test")
# Resample
dataset = dataset.cast_column("audio", Audio(sampling_rate=16_000))

processor = AutoProcessor.from_pretrained("Respeecher/ukrainian-data2vec-asr")
model = Data2VecAudioForCTC.from_pretrained("Respeecher/ukrainian-data2vec-asr")
model.eval()

sampling_rate = dataset.features["audio"].sampling_rate
inputs = processor(dataset[1]["audio"]["array"], sampling_rate=sampling_rate, return_tensors="pt")
with torch.no_grad():
    logits = model(**inputs).logits
predicted_ids = torch.argmax(logits, dim=-1)

transcription = processor.batch_decode(predicted_ids)
transcription[0]

Training Details

Training code and instructions are available on our github