ํ”ผ๋“œ๋กœ ๋Œ์•„๊ฐ€๊ธฐ
Stop Ignoring Your Snore: Building an AI Sleep Apnea Detector with Faster-Whisper and DFT ๐Ÿ’ค๐Ÿš€
Dev.toDev.to
AI/ML

Faster-Whisper์™€ DFT ๊ฒฐํ•ฉ์„ ํ†ตํ•œ ๊ณ ์ •๋ฐ€ ์ˆ˜๋ฉด ๋ฌดํ˜ธํก์ฆ ๊ฐ์ง€ ํŒŒ์ดํ”„๋ผ์ธ ๊ตฌ์ถ•

Stop Ignoring Your Snore: Building an AI Sleep Apnea Detector with Faster-Whisper and DFT ๐Ÿ’ค๐Ÿš€

wellallyTech2026๋…„ 4์›” 19์ผ5๋ถ„intermediate

AI ์š”์•ฝ

Context

์ „ํ†ต์ ์ธ ์ˆ˜๋ฉด ๋ฌดํ˜ธํก์ฆ ์ง„๋‹จ์€ ๊ณ ๋น„์šฉ์˜ ์ „๋ฌธ ์˜๋ฃŒ ์žฅ๋น„ ์˜์กด๋„๊ฐ€ ๋†’์•„ ์ ‘๊ทผ์„ฑ์ด ๋‚ฎ์Œ. ์Šค๋งˆํŠธํฐ ๊ธฐ๋ฐ˜์˜ ๊ฐ„์ด ์Šคํฌ๋ฆฌ๋‹์„ ์œ„ํ•ด ๋‹จ์ˆœ ์Œ์„ฑ ์ธ์‹์„ ๋„˜์–ด์„  ์†Œ๋ฆฌ์˜ ์งˆ๊ฐ๊ณผ ์ฃผํŒŒ์ˆ˜ ํŠน์„ฑ ๋ถ„์„์ด ์š”๊ตฌ๋จ.

Technical Solution

  • Faster-Whisper์˜ VAD(Voice Activity Detection) ๊ธฐ๋Šฅ์„ ํ™œ์šฉํ•œ ์ฝ”๊ณจ์ด ์ด๋ฒคํŠธ์˜ ์ •๋ฐ€ํ•œ Temporal Segmentation ๊ตฌํ˜„
  • DFT(Discrete Fourier Transform) ๋ฐ STFT ๊ธฐ๋ฐ˜์˜ Power Spectral Density ๋ถ„์„์„ ํ†ตํ•œ ๋ฌดํ˜ธํก ์ข…๋ฃŒ ์‹œ ๋ฐœ์ƒํ•˜๋Š” ๊ณ ์ฃผํŒŒ Gasp ๊ตฌ๊ฐ„ ์‹๋ณ„
  • Librosa๋ฅผ ์ด์šฉํ•œ Spectral Centroid ์ถ”์ถœ๋กœ ์†Œ๋ฆฌ์˜ '๋ฌด๊ฒŒ ์ค‘์‹ฌ'์„ ํŒŒ์•…ํ•˜์—ฌ ์ผ๋ฐ˜ ์†Œ์Œ๊ณผ ์ฝ”๊ณจ์ด ํŒจํ„ด์˜ ์ฃผํŒŒ์ˆ˜ ํŠน์„ฑ ๋ถ„๋ฆฌ
  • LSTM ๊ธฐ๋ฐ˜์˜ PyTorch Classifier๋ฅผ ์„ค๊ณ„ํ•˜์—ฌ MFCC, Spectral Centroid, Silence Duration์„ ํ†ตํ•ฉ ์ž…๋ ฅ๊ฐ’์œผ๋กœ ์‚ฌ์šฉํ•˜๋Š” ์‹œ๊ณ„์—ด ๋ถ„์„ ๊ตฌ์กฐ ์ฑ„ํƒ
  • CUDA ๊ฐ€์† ๋ฐ float16 ์–‘์žํ™”๋ฅผ ์ ์šฉํ•œ Faster-Whisper ๋ชจ๋ธ ๋ฐฐํฌ๋กœ ์ถ”๋ก  ์†๋„ ์ตœ์ ํ™”
  • CTranslate2 ์˜์กด์„ฑ ๋ฐ NVIDIA ๋“œ๋ผ์ด๋ฒ„ ํ™˜๊ฒฝ ์ผ๊ด€์„ฑ ํ™•๋ณด๋ฅผ ์œ„ํ•œ Docker ์ปจํ…Œ์ด๋„ˆ ๊ธฐ๋ฐ˜์˜ ๋ฐฐํฌ ์ „๋žต ์ˆ˜๋ฆฝ

์‹ค์ฒœ ํฌ์ธํŠธ

1. ์‹œ๊ณ„์—ด ์˜ค๋””์˜ค ๋ถ„์„ ์‹œ VAD๋กœ ์œ ํšจ ๊ตฌ๊ฐ„์„ ๋จผ์ € ๋ถ„๋ฆฌํ•˜์—ฌ ๋ชจ๋ธ์˜ ์—ฐ์‚ฐ ๋‚ญ๋น„๋ฅผ ๋ฐฉ์ง€ํ–ˆ๋Š”๊ฐ€

2. ์‹œ๊ฐ„ ๋„๋ฉ”์ธ(Temporal)๊ณผ ์ฃผํŒŒ์ˆ˜ ๋„๋ฉ”์ธ(Frequency) ํŠน์ง•๋Ÿ‰์„ ๊ฒฐํ•ฉํ•˜์—ฌ ๋ฐ์ดํ„ฐ์˜ ๋‹ค๊ฐ์  ๋ถ„์„์„ ์ˆ˜ํ–‰ํ–ˆ๋Š”๊ฐ€

3. ์˜๋ฃŒ/ํ—ฌ์Šค์ผ€์–ด ๋ฐ์ดํ„ฐ์˜ ํŠน์„ฑ์ƒ ๋…ธ์ด์ฆˆ ๊ฐ•๊ฑด์„ฑ ํ™•๋ณด๋ฅผ ์œ„ํ•ด ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•(Augmentation) ์ „๋žต์„ ์ˆ˜๋ฆฝํ–ˆ๋Š”๊ฐ€

4. GPU ๊ฐ€์† ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ์˜ ๋ณต์žกํ•œ ์˜์กด์„ฑ ํ•ด๊ฒฐ์„ ์œ„ํ•ด ์ปจํ…Œ์ด๋„ˆํ™”๋œ ๋Ÿฐํƒ€์ž„ ํ™˜๊ฒฝ์„ ๊ตฌ์ถ•ํ–ˆ๋Š”๊ฐ€

ํƒœ๊ทธ

#PyTorch#DFT#Signal Processing#Faster-Whisper#VAD
์›๋ฌธ ์ฝ๊ธฐ