Embedded training means that the data are transcribed, but that we don’t know the time alignment at the model or state levels.

Important material on efficiently computing the combined likelihood of the acoustic model multiplied by the probability of the language model.

Simply mentions the lexicon and language model and refers the reader to other chapters.

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.

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.

We can just use raw counts to estimate probabilities directly.

The frequency of occurrence of each N-gram in a training corpus is used to estimate its probability.

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