To perform speech recognition with HMMs involves calculating the likelihood that each model emitted the observed speech. You can skip 9.4.1 Vector Quantization.

Mel-frequency Cepstral Co-efficients are a widely-used feature with HMM acoustic models. They are a classic example of feature engineering: manipulating the extracted features to suit the properties and limitations of the statistical model.

Introduces some notation and the basic concepts of HMMs.

The difficulty of ASR depends on factors including vocabulary size, within- and across-speaker variability (including speaking style), and channel and environmental noise.

Read up to the end of 8.5 carefully. Try to read 8.6 as part of Module 7, but rest assured we will go over the concept of dynamic programming again in Module 9. We recommend you should skim 8.7 and 8.8 because the same general concepts carry forward into Hidden Markov Models (again, we’ll come back to this in Module 9). You don’t need to read 8.9 onwards. Methods like DTW are rarely used now in state of the art systems, but are a good way to start understanding some core ideas.

A simple way to generate a waveform is by concatenating speech units from a pre-recorded database. The database contains one recording of each required speech unit.

Beyond getting the phones right, we also need to consider other aspects of speech such as intonation and pausing.

Each word in the normalised text needs a pronunciation. Most words will be found in the dictionary, but for the remainder we must predict pronunciation from spelling.

We need to normalise the input text so that it contains a sequence of pronounceable words.