Natural Language Processing · 免费预览
语音信号与特征
口语如何转化为数值——从音素与发音，经过声波、频谱和共振峰，直到对数梅尔与 MFCC 特征。
1. Speech Signals and Features
Speech Signals and Features
2. Waves, sampling, and pitch
Waves, sampling, and pitch
3. Spectra, formants, and the source-filter model
Spectra, formants, and the source-filter model
4. Log mel features and MFCCs
Log mel features and MFCCs
5. Classifying a consonant
The phone [s], as at the start of sound, is produced with the tongue near the alveolar ridge forcing air through a narrow groove. What is its manner of articulation?
6. Finding the Nyquist limit
A speech waveform is sampled at Hz. Every cycle of any component wave needs at least two samples, so what is the highest frequency in Hz the sampled signal can represent?
7. Converting to the mel scale
Using the mel transform , convert a component at Hz to mels, rounding to the nearest whole mel.
8. What the cepstrum buys
MFCC computation takes the discrete cosine transform of the log mel vector. What does this step achieve that the log mel vector alone does not?
9. Counting MFCC dimensions
A classic MFCC vector keeps 12 cepstral coefficients plus one energy term, then appends the delta and the double-delta versions of all 13 values. What is the total dimensionality of the feature?