Trajectory meeting at CSTR

	Trajectory models for speech processing
	Centre for Speech Technology Research University of Edinburgh Wednesday 31st August 2005

Motivation for this meeting

Over the last few years, many different attempts have been made to overcome the limitations of the HMM. In many cases, it has been proposed that we should model trajectories; this is usually motivated by knowledge of speech production. This meeting is an opportunity to bring together some of this work and perhaps make new connections between, and interpretations of, these types of models. The main talk of the day will be by Keiichi Tokuda of the Nagoya Institute of Technology, the creator of the "Trajectory HMM".

Programme

Talks scheduled for 30 minutes may be shorter, to leave more discussion time.

Time Event Details

10.00-10.30 Coffee

10.30 Welcome Simon King

10.30-11.30 Talk Keiichi Tokuda "The Trajectory HMM"

11.30-12.00 Talk Chris Williams "How to pretend that correlated variables are independent by using difference observations"

12.00-12.30 Talk John Bridle "Interpretations of HMMs with dynamic 'observations'"

12.30-13.00 Discussion

13.00-14.00 Lunch Buffet lunch

14.00-14.30 Talk Martin Russell: new ASR, modelling and speaker recognition results from the multiple-layer HMM.

14.30-15.00 Talk Mark Gales "Augmented statistical models."

15.00-16.00 Coffee and discussion

16.00-17.00 Short/Lightning talks All participants are invited to contribute - talks may be well prepared or completely spontaneous. Suggested topics:

Khe Chai Sim: A discriminative semi-parameteric trajectory model,
Veena Singampalli: Statistical models to relate speech gestures to meaning
Hervé Bourlard: Why trajectory-based modeling (as currently proposed) will not work for speech recognition
Zhang Le: Learning a Delta Chain
Plus

Responses to earlier talks
Making connections between different models
Future research directions

17.00 Closing

Time	Event	Details
10.00-10.30	Coffee
10.30	Welcome	Simon King
10.30-11.30	Talk	Keiichi Tokuda "The Trajectory HMM"
11.30-12.00	Talk	Chris Williams "How to pretend that correlated variables are independent by using difference observations"
12.00-12.30	Talk	John Bridle "Interpretations of HMMs with dynamic 'observations'"
12.30-13.00	Discussion
13.00-14.00	Lunch	Buffet lunch
14.00-14.30	Talk	Martin Russell: new ASR, modelling and speaker recognition results from the multiple-layer HMM.
14.30-15.00	Talk	Mark Gales "Augmented statistical models."
15.00-16.00	Coffee and discussion
16.00-17.00	Short/Lightning talks	All participants are invited to contribute - talks may be well prepared or completely spontaneous. Suggested topics: Khe Chai Sim: A discriminative semi-parameteric trajectory model, Veena Singampalli: Statistical models to relate speech gestures to meaning Hervé Bourlard: Why trajectory-based modeling (as currently proposed) will not work for speech recognition Zhang Le: Learning a Delta Chain Plus Responses to earlier talks Making connections between different models Future research directions
17.00	Closing

Logistics

Registration

Please email Simon King to register - there is no charge, but we need to know how many to cater for. If you will come for dinner in the evening, please let me know.

Location

The Conference Suite at:

2, Buccleuch Place
Edinburgh EH8 9LW

Come to 2, Buccleuch (/b@'klu:/) Place and up to the first floor; ring the bell on the right hand door, then follow signs to the conference suite.

Maps can be found at the foot of this page.

Accommodation

We advise booking early, since Edinburgh is very busy in August. The best value accommodation is in University halls of residence rooms:

Edinburgh First +44 (0)131 651 2007
18 Holyrood Park Road, Edinburgh, EH16 2189

3*** accommodation at Chancellors Court - £66 per night bed and breakfast,
2** accommodation at Holland House - £44 per night bed and breakfast.
There are also a number of cheaper rooms available with shared facilities (cost is around £29).
Bookings can be made on line through the website.

For a list of hotel and guest house suggestions, see here (note - the block-booked rooms are only for the CSTR-21 event and not for this workshop)

Travel

By rail: to Edinburgh Waverley; we are a 20 minute walk from the station.

By air: it is a 25 minute bus ride (airport bus number 100) to Waverley station, then a 20 minute walk. Alternatively, a taxi direct to Buccleuch Place will cost around £15.

Participants

Provisional list, in no particular order:

Keiichi Tokuda 
Simon King 
Joe Frankel
Zhang Le
Steve Renals
Korin Richmond
John Bridle
Chris Williams
Miles Osborne
Andre Coy
Roger Moore
Vicky Maier
Neil Lawrence
Mahesan Niranjan
Martin Russell
Xiaoyan Zheng
Mark Gales
Martin Layton
Khe Chai Sim
Veena Singampalli
Thomas Hain
Esmeralda Uraga
Hywel Richards
Yasser Abdel-Haleem
Paul Taylor
Ning Ma
Richard Rohwer
Hervé Bourlard
Yoshi Shiga

Trajectory models for speech processing