T0 and SPERT-II Applications

We believe that vector microprocessors show promise in providing high performance at low cost for a wide range of applications. A number of groups are evaluating the T0 vector microprocessor and the SPERT-II workstation accelerator board. The SPERT-II board is already in production use speeding research in speech and image recognition.






Backpropagation Training of Neural Networks for Speech Recognition

David Johnson has written the QuickNet program for SPERT-II, and this is in production use training large neural networks for speech recognition research at ICSI. We have shipped SPERT-II boards to a number of our collaborators in speech research around the world to run this and related applications.


Paola Moretto investigated the implementation of the RASTA speech recognition front-end signal processing algorithm on T0.

HMM Decoding Algorithms

Su-Lin Wu has begun to look at vectorizing HMM decoding algorithms.

Kohonen Self-Organizing Feature Maps for Speech Coding

Krste Asanovic has implemented Kohonen maps on T0 for a speech coding application.

Additive Audio Synthesis

Todd Hodes is developing an additive audio synthesis engine for T0 based on the requirements of synthesis systems developed at The Center for New Music and Audio Technologies (CNMAT).

Image Processing for Traffic Surveillance and Vehicle Navigation

Chris Bregler has implemented some core image processing routines on T0.

Visual Acoustic Speech Recognition (Computer Lipreading)

Fed up with waiting for his workstation to complete image convolution tasks, Chris Bregler has also been putting SPERT-II to production use to speed his research in computer lipreading.

MPEG Decompression

Arno Formella started to look at implementing some MPEG decompression functions on T0.

Priority Encoding Transmission (PET)

PET is a technique that will recover the most important information in a message no matter which packets are lost or intentionally deleted during transmission. PET has been shown to greatly improve the performance of MPEG transmission over the Internet. Work is underway to port the PET algorithm to T0.

MPEG Encoding

Cedric Krumbein has ported the Berkeley MPEG Encoder to T0 and has been investigating vectorized MPEG motion vector estimation.


Warner Warren has been writing efficient Fast Fourier Transform code for T0.
David Johnson <davidj@ICSI.Berkeley.EDU>
$Date: 2002/02/25 20:24:17 $