Language Models: Native
speakers of different languages normally pronounce corresponding
sounds differently. For this reason, ANSR needs to be trained on
each different language.
ANSR has pronunciation models trained for Spanish, German,
Japanese, UK English, and US English. The demonstration system is
using US English models, so if you are not a native speaker of
American English, you can unfortunately expect poor recognition
results from the demo; however, excellent results should be
obtainable with language-appropriate models, which we can provide
for you, or you can build yourself.
Sprex customers are able to develop models for other languages
using the HTK toolkit from Cambridge University and can plug them
into ANSR applications.
For this reason, customers like SINTEF, in Norway, purchase ANSR:
they can now develop Norwegian speech recognition systems for
networked applications, with HTK based acoustic models used in
ANSR's action-oriented, network-distributed speech application
architecture.