Speech Coding Using an Analysis-by-Synthesis Sinusoidal Model
Signal Compression Laboratory Research Project
| Researcher: | Cagri Etemoglu |
| Faculty: | Dr. Vladimir Cuperman |
| Research Focus: | To achieve toll quality speech at
4kbps, we propose an AbS sinusoidal speech coder in which sinusoidal components are
extracted in a closed loop fashion. The problem in open loop analysis is the fact that
analysis does not correlate well with the synthesis which results in degraded speech
quality. In the closed loop method, sinusoidal components are extracted recursively.
Consequently, at each instant the encoder knows the current error, which helps it to
correlate well with the decoder. The same method can be used to analyze different types of
frames. The parameters are the amplitudes, frequencies and phases of the sinusoids.
Amplitudes will be quantized with variable dimension VQ techniques. Phases are modeled in
voiced frames as quadratic phase, in unvoiced frames as random phase and in transition
frames they are quantized. Frequencies extracted by the analysis are constrained to a
candidate frequency space so that the resulting frequencies will be amenable to efficient
modeling.
|
| Presentation: | Speech Coding Using an Analysis-by-Synthesis Sinusoidal Model |