Technology used in Melisma AI.
Melisma is a general-purpose music synthesizer using generating AI. Read the slar, articulation, and lyrics of the sheet music, and synthesize them according to the characteristics of the instrument and singing voice.
In the case of violin, how to move fingers and bows, for wind instruments, how to breathe and fingers, and for songs, the connection of sound and how to move the mouth are speculated on the neural network. By combining a normal neural network and a diffusion model, it is possible to synthesize highly expressive voice.
Because it can output different nuance performances from the same sheet music, it can be used as a retake or as an ensemble.
About learning data
The learning data uses sheet music created by professional composers and arrangers for this AI and is recorded in the studio by professional performers.
Rights such as copyrights are completely clean, and you can use them with confidence. In addition, because we use data specialized in AI, the quality of the generated audio is also high.
Distribution of revenue to players
We plan to build a system to return the support money received from everyone to the players, etc. I support sustainable music activities.
By 2025, a system will also be established that allows players to easily request performances. Our goal is to lower the threshold for ordering live music and to be able to freely choose between AI and humans.
About the name
Melisma is a singing method that extends a single syllable across several sounds. I named it so that it would be possible to connect the slar, which is difficult with conventional PCM sound sources, and to be able to express music that is impossible for humans.
Functions you want to implement in the future, etc.
・Enrichment of musical instruments and types of voices
・Improvement of generation speed and reduction of noise
・Development of a desktop application that can be edited with a piano roll
・A function that automatically attaches slar and articulation from MIDI betta
・Other
Regarding the proposal of new features
Please do not send suggestions for new features. This is because there is a risk of controversy over the right to ideas. To avoid this risk, Melisma developers will not accept any suggestions from users except when requested by developers. If you send it, we will delete it without reading it at all.
Thank you for your understanding and cooperation.
Contact information
Regarding Melisma, serious defects and media coverage, etc.
Melisma@kagura-music.jp
Please contact us.
We have a dedicated Discord server to report minor defects and bugs, so please use it.
Click here for the personal website of the producer of Melisma. If you are interested, please take a look.
License
Melisma-core Developed by Nakasako Sakana, KAGURA Music Pub., Japan
Vocoder Developed by Reo Yoneyama
@ Nagoya University, Japan
https://kagura-music.jp/melisma-technical
这东西要是能搞出更逼真的吉他、二胡之类的得多爽啊！我感觉传统音源库无论采样还是建模要有压力了。
本帖最后由 allentsei 于 25-2-8 08:57 编辑