Sixth International Conference on Spoken Language Processing
(ICSLP 2000)

Beijing, China
October 16-20, 2000

Reduced Traceback Matrix Storage for Small Footprint Model Alignment

Jeff Meunier

Human Interface Laboratory, Motorola Labs, Schaumburg, IL, USA

This paper describes two alternative techniques for storage of path traceback information during model alignment. By taking advantage of the restricted state transitions in the standard leftright hidden Markov model (HMM), traceback information can be stored as a set of state dwell counts. Compared to the traditional method of storing intermediate path information for every frame, these in-place techniques can provide significant memory savings in a small footprint embedded system. The first technique organizes the dwell counts in a "triangular array" and specifies a set of update operations performed during alignment. This direct storage technique allows for easy access of the path information once alignment is complete. The second technique encodes the dwell counts in base-D numeric values. By doing this, the array update procedure reduces to a single calculation and does not require multiple memory copies. Both techniques can be done in-place and require around N2/2 memory words instead of the NT memory words required for conventional storage (where N is the number of HMM states and T is the maximum number of frames in a modeled utterance).

Full Paper

Bibliographic reference.  Meunier, Jeff (2000): "Reduced traceback matrix storage for small footprint model alignment", In ICSLP-2000, vol.2, 835-838.