Recommendations: Language models and search

The recommendations given here apply more to research prototype  systems than real life applications:

1. For tasks using small vocabularies as in voice commands, the use of N-best method is sufficient.
2. For more complex tasks with large vocabularies , word graphs  offer many advantages. In particular, they allow the testing  of different language models without the need to rerun a full acoustic recognition.
3. For speech recognition  in the context of understanding and dialogue systems (see Chapter 13),  there is an obvious need for word graphs . Without using word graphs, it would still be possible to do the recognition job, but by using word graphs , it is much simpler to achieve an efficient algorithm for the interaction between acoustic recognition, language modelling and understanding.