Please note that these pages are no longer up-to-date.
The activities are now carried on in the HLT research unit
(hlt.fbk.eu)
DITELO
DIalogo
su TELefonO |
 |
We build speech recognizers
for telephonic applications, with the main aim to include them in call-center
systems (but not only). This means building acoustic and language models,
and offering dialogue capabilities. |
|
At present we have a speech recognizer
(have a look
at the recording of a demo) with the following features:
|
speaker independent (acoustic models are trained on several hundreds of
speakers having good geographical coverage)
Italian language
continuous speech (no pauses between words)
mixed initiative dialogue for limited domains (the user can provide information
not explicitly asked by the system)
recursive transition networks used to model the language (including bigrams,
regular expressions, etc.)
|
|
We are presently working on:
|
-
barge-in capabilities (the user can speak during the vocal prompts)
-
dialogues on complex tasks (tourism)
-
verification / confidence scores for sentences and words
-
standard interfaces (C++ API, JavaSpeech)
-
WEB access by voice (XML, VoxML languages)
-
multimodal interaction
|
|
|