Two models can both earn an overall MOS of 3.8 and still fail in very different ways. One may sound fluent and natural while regularly mispronouncing certain words. Another may pronounce every word correctly yet sound flat and mechanical. The aggregate score places those problems on the same scale, but the sub-scores separate them and point toward different fixes.
This post explains how to interpret Podonos's three sub-scores, prosody, naturalness, and pronunciation, alongside the aggregate score. It focuses on a frequent diagnostic pattern: high naturalness paired with low prosody, and what that combination generally suggests.
What the sub-scores measure
Naturalness reflects whether speech sounds like a real person talking comfortably and without strain. Its main perceptual cues include fluency, freedom from clicks and discontinuities, and audio quality that feels human-produced rather than synthesized. It responds most strongly to vocoder quality, waveform distortion, and signal problems unrelated to prosody.
Prosody measures whether rhythm, stress, intonation, and phrasing fit the text's linguistic content. Examples include a sentence-final rise on a declarative statement, stress on the wrong syllable in a polysyllabic word, or a question-like contour applied to a statement. Prosody is linguistic and structural, while naturalness concerns perceived audio quality. The two relate, but they measure different things.
Pronunciation concerns phoneme-level correctness, including whether individual phonemes and their sequences are realized as intended. It exposes errors such as /th/ becoming /d/, or a foreign-origin word receiving a rough phonetic rendering instead of its conventional English form. These mistakes are usually more localized and easier to spot than prosody failures.
The dimensions overlap without being interchangeable. Deterministic or rule-based systems may pronounce words accurately yet perform poorly on prosody. Conversely, vocoder-heavy systems may sound smooth and human-like while scoring low on pronunciation because their phonemic output is imprecise.
High naturalness, low prosody
A particularly useful split is strong naturalness, above 4.0, combined with clearly weaker prosody, below 3.5. This pattern more often suggests a training-data coverage gap than a problem with the architecture.
Naturalness responds to waveform quality, fluency, and a lack of artifacts. Once a model has enough data to generate clean audio, those qualities tend to hold across sentence types. Prosody instead asks whether stress, rhythm, and intonation fit the particular text. It therefore depends much more on whether the corpus contains the phonetic and syntactic patterns represented in the evaluation sentences.
High naturalness with low prosody usually means the audio pipeline is producing clean speech, but the model has not encountered enough similar sentence types to learn their appropriate prosodic patterns. The likely issue is coverage rather than architecture.
Which gaps can create this result? Long sentences with layered subordinate clauses may use distinctive intonation that takes many examples to learn. Questions, especially tag and indirect questions, also depend heavily on structure and may be underrepresented in read-speech corpora. Proper nouns, including company, place, and person names, form another area of variation. A model trained mainly on news text may give a proper noun different prosody from one trained on a broader corpus, and neither may fit the expectations of a speaker from another regional background.
High pronunciation, low naturalness
The reverse combination, high pronunciation with low naturalness, arises in other architectural settings and calls for a different diagnosis.
A system can produce accurate phoneme sequences while lacking the fluency of natural speech, a pattern common in parametric or partly rule-based designs. It can appear in FastSpeech-family architectures trained without a well-tuned duration predictor, or in older unit-selection systems where concatenation boundaries remain audible.
The remedy usually lies in the generation pipeline, not in broader data coverage. Inspect the vocoder, duration model, pitch prediction model, or post-processing chain. More training data rarely corrects a naturalness shortfall caused by architecture. That is why divergence between sub-scores is valuable for routing work: the visible symptom is "it sounds bad," but the responsible component depends on which score is low.
Route debugging with sub-scores
Before per-dimension results were available, diagnosing a MOS regression meant arranging a listening session. Sub-scores allow earlier triage. If prosody falls while naturalness stays steady, inspect text analysis or the duration model before the vocoder. If naturalness falls while pronunciation and prosody remain stable, investigate the waveform path, such as a vocoder checkpoint that failed to converge cleanly or preprocessing that added noise to the training signal.
Sub-score divergence is evidence for investigation, not a complete diagnosis. High naturalness with low prosody does not always indicate a training-data coverage gap. For example, an evaluation set may contain too many examples of one sentence type. The scores suggest where to look, but further inspection is needed before treating them as a root cause.
Follow sub-scores across training
Changes across consecutive checkpoints can reveal more than one isolated result. If prosody declines gradually across training runs while naturalness and pronunciation remain stable, the trend may persist unless corrected. A typical next step is to review the data for sentence types with weak prosody and add targeted examples of those patterns.
Time-series tracking can reveal a naturalness peak: a model reaches its highest naturalness at one training step, then later data improves pronunciation while naturalness begins to fall. This may indicate overfitting to a narrow prosody distribution, or a vocoder pushed beyond its comfortable range as the acoustic model improves. Without dimension-level tracking, the turn is hidden in the aggregate MOS curve, and the decline may go unnoticed until a pre-release listening session.
The aggregate score is still the right single-number summary of voice quality. But when training changes require a decision about where to investigate, the sub-scores provide the starting point.