In many tonal languages, the length of a syllable helps signal which tone it carries, alongside its pitch. "Checked" tones, found on syllables that end in a stop consonant, are reliably shorter than other tones. Speaking rate blurs this cue: a fast speaker's long tone can be as short as a slow speaker's short one. The standard fix is to divide each syllable's duration by some local measure of rate. The fix assumes that speaking faster shrinks every tone by the same factor. Precision physics has a hard-earned lesson about this kind of fix. Taking a ratio cancels only the part of an error that is shared between numerator and denominator, and whatever is not shared becomes a residual systematic uncertainty: a bias that more data will not remove. This note proposes to measure that residual for duration-based tone cues, then build a tone classifier that carries it explicitly, instead of pretending the ratio removed it.
The core claim is that tone categories respond to speaking rate with different slopes. The phonetics literature already hints at this: short syllables distort pitch shape in Taiwan Mandarin, and coda weakening in Taiwanese lengthens checked vowels. If the claim holds, rate-ratio normalization leaves a tone-dependent bias that grows with how far a speaker's rate departs from the training distribution. A classifier that treats local rate as a nuisance parameter should stay accurate and well calibrated on unusually fast or slow speech, where ratio normalization degrades. A nuisance parameter is an unknown quantity that affects the data but is not of interest, and it is estimated with its own uncertainty. If the slopes turn out to be equal, the ratio is sufficient, and that is also a clean, publishable answer.
Where the idea comes from
Two physics results supply the method. Short-baseline reactor analyses divide antineutrino spectra measured at two distances so that the unknown absolute reactor flux cancels, which lets them claim oscillations at 3.7σ without trusting the flux model (Model-independent $\bar{\nu}_e$ short-baseline oscillations from reactor spectral ratios). That cancellation works because the flux error is common to both detectors. Daya Bay's hydrogen-capture measurement of $\theta_{13}$ went further (Independent measurement of the neutrino mixing angle $\theta_{13}$ via neutron capture on hydrogen at Daya Bay). It used a second channel whose energy-scale and efficiency errors differ from the primary channel's, and it budgeted those errors separately rather than assuming them away. We borrow two things from this work: the discipline of asking which error components are common and which are not, and the use of a separately measured "control" quantity to constrain the nuisance.
The phonetics side supplies the motivation and the constraints. In Malipo Kim Mun Mien, duration is a main effect that separates checked from unchecked tones independently of pitch level (Tonal Contrasts in the Malipo Variety of the Mienic Language). This is the motivating evidence that duration carries tone identity. Functional linear models predict Taiwan Mandarin pitch contours from log-normalized duration with errors often below a semitone, but they break down at the shortest durations, where speakers hit physiological limits (Predicting taiwan Mandarin tone shapes from their duration). That breakdown is our first evidence that rate effects are not a clean common scale factor.
In Taiwanese Min Nan, weakening of coda stops lengthens checked vowels and shifts voice quality (Taiwan Min Nan (Taiwanese) Checked Tones Sound Change). This means checked tones may scale with rate differently from unchecked ones, which is exactly the non-cancelling component we want to measure. Singleton–geminate work in Libyan Arabic finds an unusually stable duration ratio of about 2.3–2.5 (Durational and Non-durational Correlates of Lexical and Derived Geminates in Arabic). That result is a warning that some ratios really do cancel, so our hypothesis may fail for some contrasts. It gives a useful reference case: a contrast where equal slopes are expected.
The closest existing system is Wang and Lee's affine-invariant tone recognizer on the COSPRO Mandarin corpus (Mandarin tone recognition using affine-invariant prosodic features and tone posteriorgram). It normalizes pitch, duration and energy under an assumed affine distortion and adds neighboring-syllable tone posteriors in a second pass, reaching 68.02% accuracy. Its most important finding for us is a negative one: normalizing variance at the syllable level hurt accuracy, because the variance held phonological information. That system chooses a normalization and commits to it as a point estimate. Our proposal differs in that it estimates how much tone-dependent rate response remains after normalization, and it propagates the uncertainty of the rate estimate into the classifier instead of subtracting a single value.
Core Claim and Model
For syllable $i$ with tone category $k$ in local context $u$, we model log duration as
$$\log D_i = \mu_k + \beta_k\, r_u + \delta_i + \epsilon_i,$$
where $\mu_k$ is the tone's typical log duration, $r_u$ is the local log speaking rate, $\beta_k$ is the tone's sensitivity to rate, $\delta_i$ is boundary-placement error from the segmenter, and $\epsilon_i$ is residual noise. Ratio normalization subtracts an estimate $\hat r_u$ built from neighboring syllables. This removes the rate term exactly only if every $\beta_k$ equals a common value $\bar\beta$. Otherwise a bias of $(\beta_k - \bar\beta)\,r_u$ remains, plus $\beta_k(r_u - \hat r_u)$ from error in the rate estimate. The bias grows linearly as rate moves away from its mean, and the second term grows as the context used to estimate rate gets shorter. These two terms are the residual systematic uncertainty we propose to measure.
Two extensions are central to the design. First, the linear-in-$r_u$ form is only a first approximation. Because the functional-model results show the duration–pitch relation bending at short durations, we let $\beta_k$ vary smoothly with $r_u$ and test whether the curvature is detectable. Second, $\delta_i$ matters for corpora whose syllable boundaries come from automatic alignment. It plays the role of the detector-calibration error in the physics analogy.
Research Program
The first stage is measurement. We fit the model above to corpora with checked and unchecked tones and to COSPRO Mandarin. We treat $r_u$ as a latent variable, constrained by a "control region": the durations of neighboring syllables, whose own tone-dependent slopes are fitted jointly with everything else. From this fit we report each tone's $\beta_k$ with its interval, the spread of $\beta_k$ across tones, and the predicted residual bias as a function of rate departure. The decisive quantity is whether the spread of $\beta_k$ is distinguishable from zero. We also report the same analysis on the Libyan Arabic geminate contrast, where the stable 2.3–2.5 ratio leads us to expect equal slopes.
To estimate $\delta_i$ on Cantonese, we hand-correct syllable boundaries on a few hundred utterances from the WenetSpeech-Yue evaluation set. That corpus's timestamps come from forced alignment (WenetSpeech-Yue: A Large-scale Cantonese Speech Corpus with Multi-dimensional Annotation). The hand-corrected subset measures boundary-error variance directly, and that variance enters the model as an explicit error term, the way an energy-scale uncertainty would.
The second stage is a classifier. We compute tone posteriors from a joint likelihood over duration and pitch-contour coefficients, with the pitch contour parameterized by intercept, slope and curvature as in the Mienic growth-curve analysis. The rate nuisance is either profiled (set to its best-fit value per hypothesis) or marginalized under its control-region uncertainty. The practical effect is simple: when an utterance is short and its rate is poorly constrained, the classifier automatically leans more on pitch and less on duration, and its confidence drops accordingly. We use the same model with neighboring tone posteriors fed back as context, so that it is directly comparable to Wang and Lee's second pass.
Baselines
The naive reference point is a classifier over the same pitch-contour coefficients and raw log duration, with no rate handling. It tells us how much rate variation costs in the first place. If this reference is barely worse than the others under rate shift, rate is not the bottleneck and the proposal is moot.
The strongest classical competitor is Wang and Lee's hybrid affine normalization plus tone posteriorgram system. It is the published method on the exact question of normalizing prosodic features for tone recognition, and it reports 68.02% five-tone accuracy on COSPRO, up from a 64.17% baseline (Mandarin tone recognition using affine-invariant prosodic features and tone posteriorgram). We will reimplement it on our COSPRO splits with the same features, the same context window and a matched hyperparameter search, and confirm that we reproduce roughly their in-distribution number before comparing. Beating it means matching it in-distribution and exceeding it on the held-out extreme-rate strata, with a better expected calibration error, which measures the gap between stated confidence and actual accuracy.
The competitor we consider most dangerous is plain local ratio normalization: duration divided by the mean duration of neighboring syllables, fed to the same classifier head. It is the direct analogue of the reactor spectral ratio, it is trivial to implement, and it is exactly what our model reduces to when all $\beta_k$ are equal. It receives an identical context window and tuning budget.
A modern learned baseline is a pretrained speech encoder fine-tuned for tone classification with speed-perturbation augmentation. Such encoders are known to be beatable on Cantonese recognition: a fine-tuned wav2vec2 reaches 15.4% character error rate on Common Voice Cantonese against 10.9% for a phonetically structured method (AMP: A unified framework boosting low resource automatic speech recognition). That is a transcription error rate, not tone accuracy, and we found no reported result on Lacuna for pretrained-encoder tone classification on COSPRO. We will therefore run this baseline ourselves on the same splits with a compute budget matched to the largest classical model's tuning.
Two ablations isolate the key ingredients. The first ties all $\beta_k$ to a single value, which tests the tone-dependent slopes. The second plugs in a point estimate of rate rather than propagating its uncertainty, which tests the error budget. If the gain survives only one ablation, we will know which ingredient carries it.
Why not just…?
Why not just divide by local syllable rate and stop? This is the real threat. If the fitted $\beta_k$ are statistically indistinguishable, ratio normalization is the right answer and our classifier collapses to it. The tied-slope ablation settles this directly. The Min Nan coda-weakening results and the Mandarin short-duration breakdown give reason to expect unequal slopes at least for checked tones and fast speech, but that is a hypothesis, not a finding. We will report a null result as such.
Why not just fine-tune a large pretrained speech model with speed perturbation? Speed perturbation stretches all segments uniformly, so it teaches the model precisely the equal-slope assumption we are testing. If the assumption is wrong, the augmentation bakes in the bias. The rate-stratified test below shows whether the fine-tuned encoder degrades on extreme-rate strata the same way ratio normalization does. If it does not degrade, the encoder has learned tone-specific rate responses from data, which would weaken the case for an explicit model, although not for the measurement.
Why not just z-score features per speaker or per syllable? Wang and Lee already found that syllable-level variance normalization lowered tone accuracy because the variance carried the contrast. Per-speaker z-scoring removes habitual rate but not rate changes within an utterance, which is the variation at issue here. We include utterance-level z-scoring inside the naive reference for completeness.
Why not rely on hidden Markov models or time warping, which already absorb timing variation? Both treat duration change as distortion to be absorbed rather than as information. The Mandarin functional-model result shows that shorter syllables change pitch-contour shape, not just its length, which uniform warping cannot represent.
Proposed Experiments and Decisive Checks
The central experiment is a rate-shift split. On COSPRO we bin utterances by estimated speaking rate, train on the middle three quintiles, and test separately on the middle and the two outer quintiles. On the Taiwanese Min Nan checked-tone corpus we split by age group and dialect region, since coda weakening differs across groups, and additionally by rate. For Cantonese checked (entering) tones we use WenetSpeech-Yue with the hand-corrected boundary subset.
The study succeeds if three conditions hold. First, the tone-slope spread is significantly nonzero in at least the checked-tone data. Second, the residual bias predicted from the measured slopes matches the accuracy loss that ratio normalization actually suffers on the outer quintiles. This is a quantitative prediction made before looking at the test results. Third, the nuisance-parameter classifier recovers a meaningful fraction of that loss with lower calibration error than all baselines, and the gain shrinks toward zero for short contexts only as fast as the rate-estimate uncertainty predicts.
The Libyan Arabic geminate contrast serves as a negative control: the method should find equal slopes there and offer no gain. If it reports gains on the geminate contrast too, the method is fitting noise, and we would treat the positive results with suspicion.
References
- Model-independent <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" altimg="si1.gif" overflow="scroll"><mml:msub><mml:mrow><mml:mover accent="true"><mml:mrow><mml:mi>ν</mml:mi></mml:mrow><mml:mrow><mml:mo stretchy="false">¯</mml:mo></mml:mrow></mml:mover></mml:mrow><mml:mrow><mml:mi>e</mml:mi></mml:mrow></mml:msub></mml:math> short-baseline oscillations from reactor spectral ratios
- Independent measurement of the neutrino mixing angle<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:msub><mml:mrow><mml:mi>θ</mml:mi></mml:mrow><mml:mrow><mml:mn>13</mml:mn></mml:mrow></mml:msub></mml:mrow></mml:math>via neutron capture on hydrogen at Daya Bay
- Tonal Contrasts in the Malipo Variety of the Mienic Language
- Predicting taiwan Mandarin tone shapes from their duration
- Taiwan Min Nan (Taiwanese) Checked Tones Sound Change
- Durational and Non-durational Correlates of Lexical and Derived Geminates in Arabic
- Mandarin tone recognition using affine-invariant prosodic features and tone posteriorgram
- WenetSpeech-Yue: A Large-Scale Cantonese Speech Corpus with Multi-dimensional Annotation
- AMP: A unified framework boosting low resource automatic speech recognition