Dolby Laboratories 585 - Manual (Page 37)

Model 585 Time Scaling Processor Principles of Operation
6-3
Frequency-domain solutions proved useful in certain situations, particularly where
processing time was not a major factor. However, they also produced characteristic
artifacts, and could only be used for a limited selection of material. Artifacts of
frequency-domain solutions often include audible phase artifacts and the loss of
definition of transient material.
6.2.3 Modeling-Based Solutions
In the 1980s, another approach was offered, this one based on signal modeling. Like
frequency-domain solutions, signal modeling could be best used on less-complex
material; different modeling applications tend to be effective on one or two types of
material only. Processing time is also significant in signal modeling, effectively
eliminating that as a real-time solution. Artifacts are still present, particularly if the
audio is not strictly in line with the profile for which a specific modeling application
was designed.
6.2.4 Choosing a Solution
Some desktop audio editing software packages include many variations of the time-
scaling solutions just described. The labels given to the choices may provide clues as
to which version may best apply to the audio you need to process, but every selection
requires a certain amount of testing before deciding what works best.
Because of the variety and complexity in some program material, this time-
consuming selection process has been, before Model 585, an exercise in choosing the
least problematic option, rather than finding an optimal solution.
Model 585 eliminates the need for such a selection. It capably handles all
combinations of speech, music, and special effects.
6.3 Intelligent Analysis
To perform time and pitch scaling, Model 585 performs a correlated multichannel
version of the time-synchronous overlap-add technique, discarding segments of the
input signal to achieve time compression, or copying and repeating segments of the
input signal to achieve time expansion. This technique is chosen as the basis of the
processing algorithm, as it is well suited to real-time application. The major
improvement Model 585 provides over existing processors is in identifying which
segments to discard or copy. To do this, Model 585 analyzes the psychoacoustic
characteristics of a multichannel audio signal in both the frequency and time domains.
To make time scaling sound as natural as possible, Model 585 identifies each
individual “auditory event” in an audio signal—such as a musical note, a drum hit, or
the phonetic components of speech. Each event is isolated and time scaled
individually. To accomplish this, Model 585 isolates audio components that have the