๊ตฌ์Œ์žฅ์•  ํ™˜์ž๋ฅผ ์œ„ํ•œ ์Œ์„ฑ์ธ์‹ ๋ชจ๋ธ

ํ”„๋กœ์ ํŠธ ์ •๋ณด

์žฌ๋‹จ๋ฒ•์ธ ๋ฏธ๋ž˜์™€ ์†Œํ”„ํŠธ์›จ์–ด์™€ ํ•จ๊ป˜ํ•˜๋Š” ์ œ 3ํšŒ์•„์ด๋””์–ด ๊ณต๋ชจ์ „

ํ”„๋กœ์ ํŠธ ๋ช…

"๊ตฌ์Œ์žฅ์•  ์Œ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ํ™œ์šฉํ•œ ๊ณ ๋ น ํ™˜์ž์˜ ์˜์‚ฌ์†Œํ†ต ๊ฐœ์„ ๋ฐฉ์•ˆ"

๋ชจ๋ธ ์„ค๋ช…

  • openai/whisper-large-v3์— ๋Œ€ํ•œ ํŒŒ์ธํŠœ๋‹ ๋ชจ๋ธ
  • ๋ณธ ๋ชจ๋ธ์€ "๊ตฌ์Œ์žฅ์•  ์Œ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ํ™œ์šฉํ•œ ๊ณ ๋ น ํ™˜์ž์˜ ์˜์‚ฌ์†Œํ†ต ๊ฐœ์„ ๋ฐฉ์•ˆ" ํ”„๋กœ์ ํŠธ์˜ ๊ตฌ์Œ์žฅ์• ํ™˜์ž๋“ค์— ๋Œ€ํ•œ ํ•œ๊ตญ์–ด ์Œ์„ฑ์ธ์‹ ๋ชจ๋ธ์ž„. OpenAI์˜ Whisper ๋ชจ๋ธ์„ ํŒŒ์ธํŠœ๋‹ ํ•˜์—ฌ ๊ตฌ์Œ์žฅ์• ์˜ ์Œ์„ฑ์  ํŠน์„ฑ์„ ๋ฐ˜์˜ํ•œ ๋ชจ๋ธ์„ ๊ตฌ์ถ•ํ•˜์˜€์Œ.
  • ์˜ค๋ฅธ์ชฝ "Inference API"๋ฅผ ํ†ตํ•ด ์Œ์„ฑ์ธ์‹ ๋ชจ๋ธ์„ ํ…Œ์ŠคํŠธ ํ•ด๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

ํ•™์Šต ๋ชจ๋ธ

  • Paper: Radford, A., Kim, J. W., Xu, T., Brockman, G., McLeavey, C., & Sutskever, I. (2023, July). Robust speech recognition via large-scale weak supervision. In International Conference on Machine Learning (pp. 28492-28518). PMLR.
  • URL: https://proceedings.mlr.press/v202/radford23a.html

ํ•™์Šต ๋ฐ์ดํ„ฐ

ํ•™์Šต ํŒŒ๋ผ๋ฏธํ„ฐ

  • learning_rate: 5e-07
  • train_batch_size: 8
  • eval_batch_size: 8
  • seed: 42
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: linear
  • lr_scheduler_warmup_steps: 10
  • mixed_precision_training: Native AMP

ํ•™์Šต ๊ฒฐ๊ณผ

Training Loss Epoch Step Validation Loss Wer
4.2932 0.09 10 4.6306 16.0442
4.2744 0.18 20 4.1942 16.2348
3.7418 0.27 30 3.7625 15.5107
3.2037 0.36 40 3.5635 14.6723
3.4714 0.45 50 3.4383 14.3674
2.8962 0.55 60 3.3494 14.1768
2.7958 0.64 70 3.2752 18.2927
2.8691 0.73 80 3.2208 19.5884
2.8693 0.82 90 3.1857 20.6174
2.9474 0.91 100 3.1644 20.6555
3.1712 1.0 110 3.1551 20.6174

Framework versions

  • Transformers 4.38.0.dev0
  • Pytorch 2.1.0+cu121
  • Datasets 2.16.1
  • Tokenizers 0.15.1
Downloads last month
38
Safetensors
Model size
242M params
Tensor type
F32
ยท
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Model tree for RecCode/whisper_final

Finetuned
(351)
this model