Speech and Audio Processing (SAP)
Speech and audio, from the physics of a sound to the models that understand it.
Instructor: Neeraj Kumar Sharma, IIT Guwahati
Modules
- 01
Seeing sound: the spectrogram
STFT, windows and resolution; reading harmonics, formants and noise in real speech.
- 02
The sounds of language
Phonemes, minimal pairs and allophones; how speech sounds are classified, and the IPA.
- 03
A buzz and a tube: the source–filter model
How speech is made: a vocal-fold source shaped by the resonances of the vocal tract.
- 04
Guessing the next sample: correlation, prediction and AR models
Correlation, linear prediction and AR models, from sunspots to speech.
- 05
Linear prediction for speech
The all-pole vocal tract: formants, LPC envelopes, analysis and resynthesis.
- 06
Speech codecs
Squeezing speech into bits: quantisation, μ-law, ADPCM and the LPC vocoder.
Updates
- Module 2 now opens with phonemes, minimal pairs, allophones and distinctive features.
- Site launched: modules 1–6 and the toolbox. New page on the spectrogram. All speech examples use the CMU ARCTIC ksp voice (Indian English) and an Indian-English recording.
Data and credits
- CMU ARCTIC speech database, speaker ksp (Indian English): Carnegie Mellon University, Language Technologies Institute; J. Kominek and A. W. Black, 2003. Used under its permissive licence; excerpts are resampled and trimmed.
- PHOIBLE 2.0: S. Moran and D. McCloy (eds.), Max Planck Institute for the Science of Human History, Jena, 2019. Data licensed CC BY 4.0.
- Glottolog: H. Hammarström, R. Forkel, M. Haspelmath and S. Bank, Max Planck Institute for Evolutionary Anthropology. Data licensed CC BY 4.0.
- TIMIT (LDC93S1) is used only for aggregate statistics such as average vowel formants; no TIMIT audio is published on this site.