NaturalSpeech: Microsoft reaches human-level quality on LJSpeech
In one sentence Microsoft Research introduces NaturalSpeech, a VAE-based end-to-end TTS system: on LJSpeech it scores -0.01 CMOS against human recordings, a difference that is not statistically significant under the Wilcoxon test (p = 0.69).
How do you measure whether an artificial voice sounds human? You play samples to real listeners and collect ratings: MOS (Mean Opinion Score) is the average rating from 1 to 5, while CMOS (Comparative MOS) measures preference by directly comparing two versions of the same sentence.
The Microsoft Research paper does two things. First, it proposes a rigorous definition of "human-level quality": a system reaches it when the difference from real recordings is not statistically significant. Second, it presents NaturalSpeech, which on LJSpeech (an English dataset of a single female voice) scores MOS 4.56 against 4.58 for the human recordings and a CMOS of -0.01.
The result holds for that dataset and those conditions: one voice, read text, good-quality recordings. It does not claim that synthetic speech is indistinguishable from human speech in every situation.
Companies
Microsoft
Tools
—
Tags
Sources