Digital Singing Voice Synthesis Using a New Alternating Reflection Model

1
Digital Singing Voice Synthesis Using a New
Alternating Reflection Model

• By Matthew Lee and Mark J. T. Smith

2
Speech vs. Singing
3
Vibrato
4
Analysis-by-Synthesis / Overlap-Add
5
The Sinusoidal Model
6
Time-Scaling with AbS/OLA
7
Problems with AbS/OLA
8
Problems with AbS/OLA
9
TD-PSOLA
10
New Model

• Vibrato Synchronous Concatenation
• Maintain continuity of vibrato by selecting
  segment boundaries according to vibrato phase.
• Alternating Reflection
• Reflect repeated portions of signal in an
  alternating manner.

11
Vibrato Synchronous Concatenation
12
Alternating Reflection
13
Alignment
14
Advantages
15
Results

• Original singing Voice
• Time-Expansion using AbS/OLA
• Time-Expansion using new Alternating Reflection
  Model

16
Future Work
17
Gaussian Model
18
Formant Tracking
19
Formant Statistics
20
Modified Vocal Tract
21
Future Work

• Identify phase conditions at boundary frames for
  optimal performance.
• Develop methods for automatic segmentation.
• Parameterize voice according to vocal quality