Lexical Tone and Word Order Matter in Translation
Prosodic Correspondence Across Languages
Center for Language and Speech Processing, Johns Hopkins University
EN I just feel like we've met before, I've seen you.
ZH 我总觉得以前好像在哪里见过你。
Data extracted from Figure 8 of the paper.
DE Ich habe immer gewusst, dass ich weiter das Mittelmeer erkunden muss.
EN I always knew I had to return to explore the Mediterranean Sea.
Data extracted from Figure 2 of the EMNLP 2026 paper.
Linking contours across languages
Pitch is extracted from both recordings, words are aligned by meaning, and each aligned cluster is compared at the same position. Explore the paper's Chinese–English example, or the German–English example from the EMNLP study.
Try it with your own voiceSay a sentence in English, Mandarin, or any language and watch its pitch contour live, including the syllable-by-syllable tones. It uses a simpler real-time method than the paper, so values will differ. Nothing is recorded or stored. Open the pitch tracker →Overview
Our EMNLP 2026 study found that pitch and energy partly carry across translation between English and three European languages. This paper asks whether that holds when languages differ in lexical tone and word order.
We pair English with Mandarin Chinese, Thai, and Vietnamese (tonal), and with Japanese and Korean (very different word order), and compare the tonal languages directly. Correspondence is lower than for European languages. Removing tone-related pitch narrows the gap but does not close it, and greater word reordering goes with weaker local correspondence. The EMNLP results serve as the published European reference throughout this page.
Findings
Correspondence is lower than for European languages
Every pair stays above its shuffled baseline, but all five English–Asian pairs fall below both the published European results and an English–French reference processed with the same pipeline.
Compare full distributions
Pick any two language pairs, including the European pairs from the EMNLP study.
Data extracted from Figure 6 of this paper and Figure 3 of the EMNLP paper; bin shares reproduce each pair's reported mean.
Lexical tone matters, but it is not the whole story
A cross-fitted model estimates the pitch pattern each tone adds to a syllable and subtracts it. Drag the slider to remove the tone component and watch the tone-specific shapes collapse toward a shared contour.
Data extracted from Figure 2 of the paper.
Does the adjustment really remove tone?
A tone classifier gets worse at guessing the tone after adjustment, though it stays above chance. Real tone labels also beat shuffled labels, so the gains come from tone information.
Word order shapes what we can measure
Japanese and Korean reorder content the most and align the least. Keeping each recording's own order (temporal pairing) raises their correlations, but most of that gain disappears once the shared sentence-level pitch decline is removed.
Detrending: a shared decline vs. word-level correspondence
"Spring nights are short" in English and Japanese. Both recordings fall in pitch from start to end.
Data extracted from Figure 3 of the paper.
More results
Which words carry the shared pitch?
Each bar is how much correspondence drops when a word class is removed. In the EMNLP study nouns mattered most for European pairs. Here, tone adjustment raises the noun contribution in all three tonal languages, close to or above French.
Timing still corresponds strongly
Utterance duration and phoneme count stay highly correlated across languages, even where pitch contours correspond weakly. Mandarin is the lowest.
Stricter semantic filtering does not change the picture
At the stricter SONAR threshold used in the EMNLP study (0.63 instead of 0.45), tonal-language correlations still stay below French.
- Alignment benchmark, Chinese
- Precision 80.5% · Recall 48.5% · AER 39.4%
- Alignment benchmark, Japanese
- Precision 71.0% · Recall 22.2% · AER 66.2%
- Contour resolution
- 10–40 frames per cluster changes means by ≤ 0.002
Takeaway
Lexical tone contributes to cross-lingual pitch differences but does not explain them alone. Local prosodic correspondence weakens between languages with very different sentence structure, even when broad sentence-level pitch patterns are shared. Expressive speech translation should model both.