| Disaggregate
Consulting |
 |
|||
|
|
|
|
|
|
| Disaggregate
Consulting |
|
|||
|
|
|
|
|
|
Disaggregate is pleased to annouce a new speech compression technology that offers remarkable fidelity, low-power operation, and the highest compression ratios of any known speech technology. "Mechanically Activated Decompression" provides superior performance using minimal system resources. This patent-pending technology may be licensed by contacting Disaggregate.
We usually model the human vocal tract as a series of pipes, chambers, and switches, as in the pictures below.
 |
 |
|||
|
Human Vocal Tract
|
Model of Human Vocal Tract
|
|||
Today's state-of-the-art speech technologies capture the model of the vocal tract in mathematical form, and use that mathematical model to compress speech, recognize speech, produce speech from text, or perform voice biometrics (e.g., recoognizing the speaker).
Mechanically Activated Decompression (MAD) introduces an entirely different appproach: a speech compression system that captures the exact utterance of the speaker and compresses it, along with a new model of the human vocal tract which allows for high-fidelity decompression.
Rather than rely on imperfect mathematical models, MAD introduces a new technology, Human Mechanical Models (HMMs). HMMs use actual moving mechanical parts to simulate the human vocal tract. The speaker talks into the HMM directly through an air tube. The HMM connects via a series of lines, pulleys, and cables (LPCs) to a mechanical recording device, while the actual utterance itself is compressed with a 1/2 horsepower air pump and placed in a standard scuba diver's tank. This latter step achieves extremly high compression ratios, with over an hour of speech in a single recording device, but at the same time recycles the original air molecules for authentic reproduction. (The operation of the upper chamber switch is controlled by a magnet; these Gaussian measurements are also recorded in more advanced versions of MAD.)
 |
 |
|||
|
Speaker Talks Into System
|
LPC System
|
|||
 |
 |
|||
|
LPC Storage Subsystem
|
Original Utterance Storage
|
|||
Playback is the opposite of compression, with the added advantage that the exact air molecules from the original utterance are available, and the air pressure of the original compression can operate the HMM to reproduce the utterance. This allows extremely low-power operation and make it especially suitable for use in systems where CPU power is at a premium, such as any Microsoft operating system.
Tests in our laboratories show that customized individual models provide superior performance over generalized HMMs. The photographs below show a typical HMM built from a human subject.
 |
 |
|||
|
Test Subject
|
HMM of Test Subject
|
|||
Disaggregate welcomes comments and questions about this exciting new technology. Demos will be available on our web site in the near future.
|
|
|
|
|
|
| Site and contents © 2001, 2002 Moshe Yudkowsky | ||||
Last updated 2002-04-23