UTMOS
A reference-free naturalness scoring model based on self-supervised speech representations that outputs a predicted MOS score from 1 to 5 for evaluating speech quality.
Core Facts
Timeline (last 90 days)
In local evaluation under the Open TTS Leaderboard protocol (not an official submission), SteerTTS v1 fp32 achieves a Seed-TTS WER of 0.96% and a UTMOS of 4.20, and SteerTTS edge achieves a Seed-TTS WER of 0.96% and a UTMOS of 4.11.
The authors state that SteerTTS edge's roughly 0.1 UTMOS loss compared with v1 comes almost entirely from the small int4 vocoder.
On the SADA corpus, the default UTMOS threshold of 3.0 would reject nearly all samples.
UTMOS is a reference-free naturalness scoring model based on self-supervised speech representations that outputs a predicted MOS score from 1 to 5.
UTMOS is trained on general English speech, and when transferred to Arabic dialects, noisy recordings or non-native accents, the score distribution shifts as a whole.
Automated machine evaluation of speech naturalness commonly uses prediction models such as UTMOS as substitutes for human evaluation.
All Facts (6)
UTMOS is a reference-free naturalness scoring model based on self-supervised speech representations that outputs a predicted MOS score from 1 to 5.
75%UnverifiedIn local evaluation under the Open TTS Leaderboard protocol (not an official submission), SteerTTS v1 fp32 achieves a Seed-TTS WER of 0.96% and a UTMOS of 4.20, and SteerTTS edge achieves a Seed-TTS WER of 0.96% and a UTMOS of 4.11.
50%UnverifiedThe authors state that SteerTTS edge's roughly 0.1 UTMOS loss compared with v1 comes almost entirely from the small int4 vocoder.
50%UnverifiedOn the SADA corpus, the default UTMOS threshold of 3.0 would reject nearly all samples.
50%UnverifiedUTMOS is trained on general English speech, and when transferred to Arabic dialects, noisy recordings or non-native accents, the score distribution shifts as a whole.
50%UnverifiedAutomated machine evaluation of speech naturalness commonly uses prediction models such as UTMOS as substitutes for human evaluation.
50%