The AT&T Statistical Dialog Toolkit (ASDT) enables developers to build spoken dialog systems
that track a distribution over multiple dialog states. The engine provided by the toolkit updates
this distribution efficiently, in real time, during the dialog. The engine is implemented in
Python. The source code of the toolkit is available under a non-commercial license.
The engine operates on a collection of dialog partitions. Each partition is a programmatic object that supports a small set of simple methods. The application developer implements the objects to suit their application.
The toolkit includes an end-to-end dialog system -- a voice dialer with a database of 100,000 (fictious) listings. The ASDT engine runs locally. You can use your own speech recognition and text-to-speech, or use those provided by the AT&T Speech Mash-ups platform. To use AT&T Speech Mash-ups, you just need to create a free account.
Download ASDT. Current version is 1.1.
To join the ASDT discussion list, email asdt-discuss-request@research.att.com with subject "subscribe". To confirm the subscription, reply to the confirmation email. We welcome questions, comments, and bug reports.
The following two videos show example interactions with the voice dialer example dialog system, which is included with ASDT. It uses AT&T Speech Mash-ups for speech recognition and TTS.
- In this video, no speech recognition errors occur. This shows a "sunny day" path through.
- In this video, there are several speech recognition errors. In the last turn, the correct hypothesis (MCCULLOUGH) is in the N=2 position on the N-Best list, illustrating how the system is able to make use of all the information on the N-Best list.
The algorithm used by the ASDT engine is described in this paper:
- Jason D. Williams. 2010. Incremental partition recombination for efficient tracking of multiple dialog states. Proc ICASSP, Dallas, Texas, USA.
- Tutorial on statistical approaches to dialog sysytems: Jason D. Williams, Steve Young, and Blaise Thomson. 2009. Statistical approaches to dialogue systems. Tutorial at Interspeech, Brighton, United Kindgom. [slides only] [slides + all movies]
- Interactive webpage that illustrates the benefit of tracking multiple dialog states.
- Bibliography of about 200 papers related to statistical dialogue