midicnAI Music 中文

Synthetic Vocals

Synthetic singing handles demos, backing and languages you cannot record; long notes and emotion still give it away.

The problem

Synthetic singing means a model performs the melody and the words, not that your own recording is shifted. It holds pitch and pronounces text well enough for three jobs: a demo someone else will sing, a backing layer under a real voice, and a language you cannot cast. It is a poor choice for the lead on a release unless listeners know the voice is synthetic.

How it works

The system separates pitch, duration and phonemes. Lyrics become phonemes, the melody supplies pitch and length, and timbre comes from recorded samples or a trained acoustic model. Separate control makes changing words, key or tempo cheap. It also means someone must decide how the line is sung: where to breathe, which syllable to weight, how a long note ends. If nobody decides, the model returns an average of those choices.

Practice

Decide the use first. A demo can be fully synthetic. On a release, replace the lead with a singer and keep the synthetic part as backing. Leave room for breaths when writing the melody, keep phrases under eight to ten syllables, and keep the range inside an octave and a half. Fade long notes in the MIDI and add a very low breath layer. Have a native speaker check stress and liaison in a foreign language. In the mix, narrow the synthetic voice slightly and clear space around 2 to 5 kHz for the real voice or dialogue.

What to listen for

In Du bist die Ruh, listen to where the long phrases break for breath. In Erlkönig, listen to how one singer colours three characters. In Monteverdi's madrigal, listen to how each line moves on its own. In the hymn, listen to how voices blend with the organ. Synthetic singing fails in exactly these places: breath, character, independence and blend.

Listen in the library

Du bist die RuhC11:31舒伯特长句歌曲,听换气与长音收尾处的能力边界(C1)DetailsDer Erlkönig, D.328C13:54舒伯特叙事歌,一人多角色,听音色转换能否成立(C1)Detailsmadrigal.3.1C13:22蒙特威尔第牧歌,多声部织体里合成声部该退到哪里(C1)DetailsO Sing to the Lord Whose Bountiful HandC10:33会众人声与管风琴,听合成和声垫在齐唱里的融合度(C1)Details

All tracks from midicn-lib — inline badge is the license tier (C1 commercial / C2 non-commercial / C3 study only).

Sources

  • 合成人声的工作方式,依公开研究摘要与产品文档核对,未抄原文
  • How synthetic singing is built, checked against public research summaries, no text copied

Used for fact-checking only; all prose is original.