A complete issue of IEEE Signal Processing Magazine. Although a few years old, this is still a very useful survey of current techniques.
Holmes & Holmes – Chapter 11 – Improving Speech Recognition Performance
We mitigate the over-simplifications of the model using ever-more-complex algorithms.
Jurafsky & Martin – Section 4.4 – Perplexity
It is possible to evaluate how good an N-gram model is without integrating it into an automatic speech recognition. We simply measure how well it predicts some unseen test data.
Jurafsky & Martin – Section 4.3 – Training and Test Sets
As we should already know: in machine learning it is essential to evaluate a model on data that it was not learned from.
Jurafsky & Martin – Section 4.2 – Simple (Unsmoothed) N-Grams
We can just use raw counts to estimate probabilities directly.
Jurafsky & Martin – Section 4.1 – Word Counting in Corpora
The frequency of occurrence of each N-gram in a training corpus is used to estimate its probability.
Jurafsky & Martin – Section 9.8 – Evaluation
In connected speech, three types of error are possible: substitutions, insertions, or deletions of words. It is usual to combine them into a single measure: Word Error Rate.
Jurafsky & Martin – Section 9.7 – Embedded training
Embedded training means that the data are transcribed, but that we don’t know the time alignment at the model or state levels.
This is the new version. Still under construction.