Lyric Intelligibility

Synthetic vocals usually fail at being understood, and the fault lies in consonants and word endings.

Related

VoicePopular singing and the microphoneLegato

The problem

The most common complaint about synthetic vocals is not that the timbre sounds fake. It is that the words cannot be understood. Once the melody gets fast or high, consonants disappear first, word endings follow, and listeners reach for subtitles. Intelligibility can be tested instead of guessed, and you do not have to wait for an audience to tell you.

How it works

Understanding a word depends on two things. Consonants mark where words begin and end; plosives and fricatives sit between roughly 2 and 8 kHz, and they are short, quiet and easily covered by an accompaniment. Vowels identify the word through formant positions. A synthesis model is trained to sound like a voice, not like this sentence, so it smooths brief consonants and merges word endings into the next note. Languages with lexical tone add another layer: the tone has to agree with the melodic direction, or the syllable is right and still sounds wrong.

Practice

Record a slow, syllable-by-syllable reference first and confirm which note each syllable lands on. Write the melody with tone in mind; a rising tone can go up, a falling tone should not be forced to a high note. Split any phrase longer than eight syllables and break it at a breath. In the mix, do not over-compress 2 to 5 kHz, keep reverb short, and set pre-delay above thirty milliseconds. Test by having three people write down the lyrics blind. One wrong word sends that line back. Then listen on a phone speaker, which exposes missing consonants faster than monitors.

What to listen for

In Mandoline, listen to French nasal vowels keeping their shape in fast syllables. In Come Away, Death, listen to how English consonants start the words. In Erlkönig, listen to one singer separating three characters by diction alone. These are the first things a synthetic voice loses.

Sources

• 语音可懂度与共振峰的关系，依语音学通识核对，未抄原文
• Consonants, vowels and intelligibility, checked against general phonetics references, no text copied

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