Prosody and pronunciation: reading beyond overall MOS
See how Podonos's naturalness, prosody, and pronunciation sub-scores expose distinct TTS failure modes and help direct debugging.
Journal
Writing on TTS evaluation
Notes on voice quality measurement, MOS methodology, and building automated listening pipelines.
TTS scoring across speaking rates
Testing TTS at 0.8x, 1.0x, and 1.4x shows how naturalness, prosody, and pronunciation diverge, revealing rate-specific model weaknesses.
Expressive TTS emotion scoring: what MOS misses
See where automated MOS helps with expressive TTS, where its limits appear, and how targeted human evaluation can fill the gaps.
MOS vs. WER: two metrics for TTS teams
Learn when MOS or WER fits TTS evaluation, where they differ, and how to combine both to assess intelligibility and voice quality.
Open-source TTS benchmark: six architectures
A comparison of 6 open-source TTS architectures using Podonos scores on the Harvard Sentence List, highlighting trade-offs across naturalness, prosody, and pronunciation.
Beyond aggregate MOS: complementary evaluation signals
Aggregate MOS can miss key signals in voice cloning, domain vocabulary, and code-switching; add speaker, pronunciation, and transition checks.
A daily voice evaluation loop without a full QA team
Set up a lightweight automated daily MOS evaluation loop for TTS teams, using fixed audio, thresholds, and clear points for human review.
How Podonos calibrates its listener model
Behind Podonos calibration: a 2400-rating listening study using ACR, recruited raters, and R-squared to explain model agreement with human scores.
API-first voice scoring across existing toolchains
Podonos chose REST instead of a Python library so teams can add voice scoring to PyTorch, Kaldi, and NeMo workflows without changing their environments.
How naturalness is judged in speech synthesis
Naturalness scores compress several perceptual dimensions; learn what they measure and why automated MOS has limited resolution at the top of the scale.
Automated MOS for TTS: a developer's guide
MOS measures perceived speech quality on a 1-5 listener scale; rooted in ITU-T P.800 from 1996, automated estimators predict ratings from TTS audio.
Meet Podonos
Why we built an automated voice quality scoring API, what problem it solves for TTS teams, and what the free tier includes.