| Paper: | SP-P13.1 | ||
| Session: | General Topics in Robust Speech Recognition | ||
| Time: | Thursday, May 20, 13:00 - 15:00 | ||
| Presentation: | Poster | ||
| Topic: | Speech Processing: Robust Speech Recognition | ||
| Title: | CODEBOOK DESIGN FOR ASR SYSTEMS USING CUSTOM ARITHMETIC UNITS | ||
| Authors: | Xiao Li; University of Washington | ||
| Jonathan Malkin; University of Washington | |||
| Jeff Bilmes; University of Washington | |||
| Abstract: | Custom arithmetic is a novel and successful technique to reduce the computation and resource utilization of ASR systems running on mobile devices. It represents all floating-point numbers by integer indices and substitutes a sequence of table lookups for all arithmetic operations. The first and crucial step in custom arithmetic design is to define and quantize system variables, preferably to low precision. In order to quantize system variables with high entropies, we explore several reordering and rescaling techniques, including linear and tree-structure accumulations in likelihood evaluation and normalization of Viterbi search. Furthermore, we investigate a discriminative distortion measure for scalar quantization to better maintain recognition accuracy. Experiments on an isolate-word recognition show that each system variable can be scalar quantized to less than 8 bits using standard quantization method, except for the alpha probability in Viterbi search which requires 10 bits. However, using our normalization and discriminative distortion measure, the forward probability can be quantized to 9 bits, thereby halving the correspoding lookup table size. This greatly reduces the system band-width and enables the implementation of custom arithmetic on ASR systems. | ||
| Back | |||
Home -||-
Organizing Committee -||-
Technical Committee -||-
Technical Program -||-
Plenaries
Paper Submission -||-
Special Sessions -||-
ITT -||-
Paper Review -||-
Exhibits -||-
Tutorials
Information -||-
Registration -||-
Travel Insurance -||-
Housing -||-
Workshops