A local virtual signer project, LINNE, is proposed several years ago. However, to process a huge amount of sound-bank data is big problem. Here we make use of the python tool lib., PyMIR and SciKit-Learn, to help us extract the necessary information that needed for a song synthesizer, ex. UTAU.
9. Sound bank
From commercial company
or volunteers
Editor
Notes and Lyrics
Vocal Synth.
Synthesized
song
Block diagram for
a VS system
Songs
10. 飴屋 P - UTAU の基本的アルゴリズムと開発経緯
http://udn.utau-synth.com/documents/kouen/20120325/
Step 1: cut the recored sounds through into sound
elements (phonons)
11. 飴屋 P - UTAU の基本的アルゴリズムと開発経緯
http://udn.utau-synth.com/documents/kouen/20120325/
Step 2: connect the elements following the lyrics
12. 飴屋 P - UTAU の基本的アルゴリズムと開発経緯
http://udn.utau-synth.com/documents/kouen/20120325/
Step 3: Adjust the pitches and lengths of the lyrics
41. Using SVM to determine
the vowel positions
http://www.cmlab.csie.ntu.edu.tw/~cyy/learning/tutorials/SVM3.pdf
https://en.wikipedia.org/wiki/Support_vector_machine
http://www.csie.ntu.edu.tw/~cjlin/libsvm/index.html
44. Supervised learning -
Training sample?
https://github.com/yuanchao/linne-analyzer/blob/vowel_det/src/linne/analyzer/cmd/linne-train2.py
45. Take part of the data as
the training sample –
Data-driven Analysis
https://github.com/yuanchao/linne-analyzer/blob/vowel_det/src/linne/analyzer/cmd/linne-spect2.py