Phonetic and semantic analyses of spoken corpora of Beijing and Taiwan Mandarin indicate that the neutral tone is a lexical tone

Technical Summary: Phonetic and Semantic Analyses of Spoken Corpora of Beijing and Taiwan Mandarin

Problem Statement
The neutral (or "floating") tone in Mandarin Chinese (T5) is traditionally described as a reduced, context-dependent tone that lacks a fixed lexical target, often appearing only on the second syllable of disyllabic words. However, its status as a distinct lexical tone remains debated. While recent studies on Taiwan Mandarin have identified word-specific pitch signatures linked to semantics for lexical tones, it is unclear if the neutral tone exhibits similar properties. Furthermore, significant dialectal differences exist between Beijing and Taiwan Mandarin regarding the prevalence and realization of the neutral tone. This study addresses four core questions:

1. Does the neutral tone possess its own distinct tonal target and pattern, or is it merely a context-dependent reduction?
2. How do the pitch contours of neutral-tone patterns differ between Beijing and Taiwan Mandarin?
3. Do disyllabic words with a neutral tone exhibit word-specific pitch signatures in both dialects?
4. Can these word-specific signatures be predicted from contextualized semantic embeddings, and do dialectal differences in realization stem from differences in word meaning?

Methodology
The study utilizes spontaneous conversational speech corpora from two varieties: Beijing Mandarin (50 speakers) and Taiwan Mandarin (55 speakers).

• Data Selection: The authors extracted disyllabic words where the first syllable carries a lexical tone (T1–T4) and the second syllable is a neutral tone (T5), based on authoritative dictionaries (Xiandai Hanyu Cidian for Beijing and the Ministry of Education Dictionary for Taiwan). A specific "Others" category was created for Taiwan data where words listed as T5 in Beijing are listed with full lexical tones in Taiwan.
• Acoustic Processing: Fundamental frequency (F0) was extracted using Praat across the entire syllable. Data were cleaned to remove tracking errors and creaky voice.
• Statistical Modeling: The authors employed Generalized Additive Mixed Models (GAMs) using the mgcv package in R. The response variable was the log-transformed F0 over normalized time. The models decomposed pitch contours into additive components tied to predictors: normalized time, tone pattern, word identity, speaker, tonal context (preceding/following tones), word position, duration, and bigram probabilities. An AR(1) process was included to account for autocorrelation in pitch time series.
• Semantic Analysis: To investigate the link between meaning and pitch, the authors used contextualized embeddings (CEs) derived from the Qwen-2.5 large language model. Word tokens were mapped to 896-dimensional vectors. A linear mapping was trained to project these semantic vectors onto the denoised, word-specific pitch signatures extracted from the GAMs. Performance was evaluated via nearest-neighbor classification accuracy against a permutation baseline.

Key Results

1. Nature of the Neutral Tone: The GAM analysis reveals that the neutral tone has a distinct, systematic tonal target (mid-to-low final pitch) that varies depending on the preceding lexical tone. Crucially, each tone pattern (e.g., T1-T5, T2-T5) exhibits a unique pitch contour component that cannot be reduced to a general sandhi rule. This suggests the neutral tone functions as a lexical tone with its own underlying patterns.
2. Dialectal Differences:
   • Beijing Mandarin: Shows an initial rise in T1-T5 and T4-T5 patterns. The T2-T5 pattern exhibits a rise.
   • Taiwan Mandarin: T1-T5 and T4-T5 lack the initial rise seen in Beijing. T2-T5 is realized as a nearly level tone. T3-T5 is the most similar across dialects, resembling a drawn-out dipping tone.
   • Variability: Taiwan Mandarin exhibits greater variability in the realization of neutral tone patterns compared to Beijing Mandarin.
3. Word-Specific Signatures: Strong evidence was found for word-specific pitch signatures in both dialects. In fact, the variable importance of "word" identity was higher in the Beijing dataset than in the Taiwan dataset. These signatures persist even after controlling for tone patterns, prosodic factors, and tonal context.
4. Semantic Predictability: Contextualized embeddings successfully predicted word-specific pitch signatures with accuracy significantly above chance (Beijing: ~24% test accuracy; Taiwan: ~20% test accuracy, against ~10% baselines). Furthermore, the "prototypical meanings" (centroids of CEs) for tone patterns predicted pitch contours that closely matched the GAM-observed contours. This suggests that dialectal differences in pitch realization may partly arise from differences in the contextualized meanings of words in the two corpora.

Significance and Claims
The paper argues that the neutral tone in both Beijing and Taiwan Mandarin should be classified as a lexical tone. It possesses a specific tonal target, exhibits systematic variation based on preceding tones, and carries word-specific pitch signatures that are partially predictable from semantic context.

The authors claim that the variability observed in the neutral tone is not unique to it but is a feature shared with other lexical tones, challenging the view of the neutral tone as merely a "reduced" or "toneless" element. The study demonstrates that the Discriminative Lexicon Model's principle—that phonetic form is linked to meaning—extends to the neutral tone. Additionally, the research highlights that dialectal differences in phonetic realization can be traced, at least in part, to differences in the semantic usage (contextualized meanings) of words in spontaneous speech, as captured by large language model embeddings.

The study concludes that while the neutral tone is restricted to the second syllable in disyllabic words, its phonetic and semantic behavior mirrors that of the four canonical lexical tones, warranting its treatment as a distinct lexical category in both varieties.