prof. seungchul lee industrial ai lab....–for short enough windows, speech can be considered to be...
TRANSCRIPT
![Page 1: Prof. Seungchul Lee Industrial AI Lab....–For short enough windows, speech can be considered to be stationary –Remember, though, that there is a time-frequency tradeoff here 10](https://reader034.vdocuments.mx/reader034/viewer/2022050301/5f6ab7448efac7728605ad9f/html5/thumbnails/1.jpg)
신호처리 (FFT and STFT)
Prof. Seungchul Lee
Industrial AI Lab.
![Page 2: Prof. Seungchul Lee Industrial AI Lab....–For short enough windows, speech can be considered to be stationary –Remember, though, that there is a time-frequency tradeoff here 10](https://reader034.vdocuments.mx/reader034/viewer/2022050301/5f6ab7448efac7728605ad9f/html5/thumbnails/2.jpg)
Signal Representation by Harmonic Sinusoids
• Question– How to decompose it to harmonic sinusoids
– How to find the coefficients of harmonic sinusoids
– How to find the frequency components
→ Fourier Transform by FFT
2
![Page 3: Prof. Seungchul Lee Industrial AI Lab....–For short enough windows, speech can be considered to be stationary –Remember, though, that there is a time-frequency tradeoff here 10](https://reader034.vdocuments.mx/reader034/viewer/2022050301/5f6ab7448efac7728605ad9f/html5/thumbnails/3.jpg)
Fast Fourier Transform (FFT)
3
![Page 4: Prof. Seungchul Lee Industrial AI Lab....–For short enough windows, speech can be considered to be stationary –Remember, though, that there is a time-frequency tradeoff here 10](https://reader034.vdocuments.mx/reader034/viewer/2022050301/5f6ab7448efac7728605ad9f/html5/thumbnails/4.jpg)
FFT with Sampling Frequency
• You can think this signal as a vibration signal from a rotating machinery with 60 Hz
4
![Page 5: Prof. Seungchul Lee Industrial AI Lab....–For short enough windows, speech can be considered to be stationary –Remember, though, that there is a time-frequency tradeoff here 10](https://reader034.vdocuments.mx/reader034/viewer/2022050301/5f6ab7448efac7728605ad9f/html5/thumbnails/5.jpg)
Implementing FFT Routine
5
![Page 6: Prof. Seungchul Lee Industrial AI Lab....–For short enough windows, speech can be considered to be stationary –Remember, though, that there is a time-frequency tradeoff here 10](https://reader034.vdocuments.mx/reader034/viewer/2022050301/5f6ab7448efac7728605ad9f/html5/thumbnails/6.jpg)
Single-sided FFT (or Positive FFT)
• Only want the second half of the FFT, since the last is redundant
• 2×amplitude except the DC component
6
![Page 7: Prof. Seungchul Lee Industrial AI Lab....–For short enough windows, speech can be considered to be stationary –Remember, though, that there is a time-frequency tradeoff here 10](https://reader034.vdocuments.mx/reader034/viewer/2022050301/5f6ab7448efac7728605ad9f/html5/thumbnails/7.jpg)
STFT (Short-Time Fourier Transformation)
7
![Page 8: Prof. Seungchul Lee Industrial AI Lab....–For short enough windows, speech can be considered to be stationary –Remember, though, that there is a time-frequency tradeoff here 10](https://reader034.vdocuments.mx/reader034/viewer/2022050301/5f6ab7448efac7728605ad9f/html5/thumbnails/8.jpg)
Issue on FFT
8
![Page 9: Prof. Seungchul Lee Industrial AI Lab....–For short enough windows, speech can be considered to be stationary –Remember, though, that there is a time-frequency tradeoff here 10](https://reader034.vdocuments.mx/reader034/viewer/2022050301/5f6ab7448efac7728605ad9f/html5/thumbnails/9.jpg)
Issue on FFT
9
![Page 10: Prof. Seungchul Lee Industrial AI Lab....–For short enough windows, speech can be considered to be stationary –Remember, though, that there is a time-frequency tradeoff here 10](https://reader034.vdocuments.mx/reader034/viewer/2022050301/5f6ab7448efac7728605ad9f/html5/thumbnails/10.jpg)
Short-Time Fourier Transformation
• The spectral content of speech changes over time (non-stationary)– As an example, formants change as a function of the spoken phonemes
– Applying the DFT over a long window does not reveal transitions in spectral content
• To avoid this issue, we apply the DFT over short periods of time– For short enough windows, speech can be considered to be stationary
– Remember, though, that there is a time-frequency tradeoff here
10
![Page 11: Prof. Seungchul Lee Industrial AI Lab....–For short enough windows, speech can be considered to be stationary –Remember, though, that there is a time-frequency tradeoff here 10](https://reader034.vdocuments.mx/reader034/viewer/2022050301/5f6ab7448efac7728605ad9f/html5/thumbnails/11.jpg)
Short-Time Fourier Transformation
• Define analysis window
• Define the amount of overlap between windows – e.g., 30%
• Define a windowing function – e.g., Hann, Gaussian
• Generate windowed segments – Multiply signal by windowing function
• Apply the FFT to each windowed segment
11
![Page 12: Prof. Seungchul Lee Industrial AI Lab....–For short enough windows, speech can be considered to be stationary –Remember, though, that there is a time-frequency tradeoff here 10](https://reader034.vdocuments.mx/reader034/viewer/2022050301/5f6ab7448efac7728605ad9f/html5/thumbnails/12.jpg)
STFT in Python
• Python function ‘spectrogram’ is useful for easily computing STFTs
12
![Page 13: Prof. Seungchul Lee Industrial AI Lab....–For short enough windows, speech can be considered to be stationary –Remember, though, that there is a time-frequency tradeoff here 10](https://reader034.vdocuments.mx/reader034/viewer/2022050301/5f6ab7448efac7728605ad9f/html5/thumbnails/13.jpg)
STFT: Time and Frequency
13
![Page 14: Prof. Seungchul Lee Industrial AI Lab....–For short enough windows, speech can be considered to be stationary –Remember, though, that there is a time-frequency tradeoff here 10](https://reader034.vdocuments.mx/reader034/viewer/2022050301/5f6ab7448efac7728605ad9f/html5/thumbnails/14.jpg)
STFT Resolution Trade-off
14