Chapter 15
Speech Features: From Moving Air to a Matrix of Numbers
Chapter: Phonetics and Speech Feature Extraction, in Daniel Jurafsky and James H. Martin, Speech and Language Processing, 3rd edition draft, August 19, 2026.
Topics to learn
Page numbers refer to the linked chapter PDF.
- Speech sounds p. 1
- Phonetic transcription p. 1
- Articulation p. 3
- Prosody: accent, stress, structure, and tune p. 8
- Acoustics p. 10
- Waves p. 10
- Frequency, amplitude, pitch, and loudness p. 12
- Reading phones from a waveform p. 15
- Spectra p. 15
- The source-filter model p. 19
- Log-mel features p. 20
- Sampling and quantization p. 21
- Aliasing and the Nyquist limit p. 21
- Windowing and frames p. 22
- The discrete Fourier transform p. 23
- Mel filters and the log p. 24
- MFCCs p. 25
- Cepstral coefficients as a compressed spectral envelope p. 25
- Why a spectrogram is not a phonetic transcription p. 15
FAQ
Does a higher bit depth prevent aliasing?
No. Bit depth quantizes amplitude. Aliasing comes from sampling too slowly for the frequencies in the signal. Filter or sample faster to control it.
Does doubling the fundamental frequency double the formants?
No. Fundamental frequency sets the harmonic spacing of the source. Formants are resonances of the vocal-tract filter. The chapter treats them as separate.
Why can a one-second recording yield fewer frames than the hop size suggests?
Each frame needs a full window, not just a hop. Without padding, the last start position is the last one that still has a complete window inside the recording.
Does zero-padding a frame add observed speech?
No. It adds zeros so the transform can use a chosen length. It samples the same finite frame on a denser frequency grid.