Phonetic-to-Articulatory Mapping

• Phonetic-to-Articulatory Feature Mapping is a systematic technique that connects phonetic representations with quantitative articulatory measures, vital for speech modeling and cross-linguistic studies.
• Static mapping using predetermined feature vectors (e.g., PanPhon and CLTS) offers clear, language-neutral phonetic encoding that supports robust speech inversion and clustering.
• Dynamic neural models capture coarticulation and speaker-independent articulatory dynamics, achieving high correlations (up to r=0.88) between predicted and measured articulator trajectories.

Phonetic-to-Articulatory Feature Mapping refers to the systematic association or predictive transformation between discrete phonetic/phonological representations (e.g., segmental symbols, feature bundles, or phone sequences) and quantitative articulatory representations (e.g., continuous trajectories of speech articulators or categorical articulatory feature vectors). This mapping underlies computational modeling of speech production, cross-linguistic sound analysis, self-supervised speech modeling, and applications such as acoustic-to-articulatory inversion, text-to-speech, and biofeedback for speech therapy. Phonetic-to-articulatory models operate across levels: hard-coded static feature encodings for cross-linguistics, dynamic sequence-to-sequence regressions with deep neural networks, and differentiable forward/inverse motor models in neural imitation systems (Singh et al., 2019, Wang et al., 2023, Azzouz et al., 4 Nov 2024, Tandazo et al., 22 Dec 2025, Rubehn et al., 7 May 2024, Lavechin et al., 6 Sep 2025).

1. Articulatory Feature Representations

Phonetic-to-articulatory mapping requires the formalization of articulatory feature spaces. These range from hand-crafted categorical feature inventories to continuous-valued geometric descriptors:

• Phonological/Articulatory Features (Categorical)
  • PanPhon system: 22 ternary-valued features (e.g., place, manner, voicing, vowel height, roundness) per IPA segment; values are {+, 0, –}; features with zero are masked from training loss (Tandazo et al., 22 Dec 2025).
  • CLTS system: 39 binary (±1/0) features covering manner, place, phonation, airstream, stridency, tone, diphthong trajectory, among others, mapped from descriptive CLTS bundles; ternary encoding distinguishes 'present', 'absent', and 'inapplicable' (Rubehn et al., 7 May 2024).
  • Ahmed et al. scheme: For vowels—continuous open (height), back (backness), binary rounded; for consonants—continuous place (lips to glottis), categorical manner, binary voiced, aspirated, pharyngeal, and airflow mode (Ahmed et al., 2020).
• Articulatory Kinematics (Continuous)
  • Electromagnetic articulography (EMA) or MRI contour points yield real-valued trajectories, e.g., 12D positions (tongue, lips, jaw sensors), or 100D tongue contours (Azzouz et al., 4 Nov 2024, Singh et al., 2019).
  • Derived vocal tract variables (TVs) such as Lip Aperture (LA), Tongue Body Constriction Location (TBCL), and others (6D), produced from pellet or geometric data (Siriwardena et al., 2022, Lavechin et al., 6 Sep 2025).

These representation choices determine the mapping's granularity, linguistic interpretability, and cross-linguistic scalability.

2. Static and Dynamic Mapping Methodologies

Static Mapping: Feature Vectors for IPA Segments

• Deterministic Mapping: IPA symbols are assigned a fixed feature vector via a lookup in curated tables, either through systems such as PanPhon or CLTS (Tandazo et al., 22 Dec 2025, Rubehn et al., 7 May 2024, Ahmed et al., 2020).
• Worked Examples: Segment t͡s in CLTS, for example, is a union of features from its subparts, such as [+coronal], [–sonorant], [+strident]. Overwrite rules establish specificity hierarchies to resolve feature redundancy (Rubehn et al., 7 May 2024).

System	#Features	Feature Values	Coverage (segments)	Example Reference
PanPhon	22	{+, 0, –}	~6,367	(Tandazo et al., 22 Dec 2025)
CLTS	39	{+1, 0, –1}	~8,700	(Rubehn et al., 7 May 2024)
Ahmed et al.	6–7	real/binary	IPA subset	(Ahmed et al., 2020)

These mappings are critical for cross-linguistic analysis, phonetic distance computations, clustering, and mapping phoneme inventories across languages.

Dynamic and Predictive Mapping: Sequence Models

• Neural Sequence Mapping: Deep bidirectional LSTMs, attention-based models, or joint-architecture systems map time-aligned phoneme sequences to articulatory trajectories (Singh et al., 2019, Wang et al., 2023, Azzouz et al., 4 Nov 2024).
• Model Inputs: Inputs may be one-hot phoneme sequences, phoneme-timing matrices, or, in contemporary systems, contextually enriched embeddings from self-supervised speech encoders (e.g., wav2vec 2.0) (Lavechin et al., 6 Sep 2025, Tandazo et al., 22 Dec 2025).
• Objective Functions: Typical loss functions include frame-level mean squared error (MSE) between predicted and ground-truth articulator positions, with optional multi-task objectives—e.g., cross-entropy for phoneme classification, enabling the network to couple articulatory and phonetic knowledge (Azzouz et al., 4 Nov 2024, Wang et al., 2023).

Dynamic mapping architectures allow frame-synchronous prediction of multiple articulators, capture coarticulatory effects, and scale to speaker-independent or multilingual settings.

3. Quantitative Metrics and Modeling Results

Evaluation of phonetic-to-articulatory mapping systems employs several quantitative criteria:

• Pearson Product-Moment Correlation ($r$): Measures the linear correspondence between predicted and measured articulator trajectories. Values for phoneme-only models range up to $r=0.81$; best joint models (acoustics + phoneme features) achieve up to $r=0.88$ (Singh et al., 2019, Wang et al., 2023).
• RMSE (mm): For MRI-based full-contour prediction, e.g., best median RMSE = 2.21 mm over 100D tongue contours (Azzouz et al., 4 Nov 2024).
• Frame-wise Feature or Phone Accuracy: In supervised representation learning, e.g., MauBERT: 95.6% articulatory feature accuracy across 55 languages (Tandazo et al., 22 Dec 2025).
• ABX Discriminability: Minimal pair discrimination tests over learned representations, used to quantify context-invariance and cross-lingual mapping quality (Tandazo et al., 22 Dec 2025, Lavechin et al., 6 Sep 2025).

Results indicate that phoneme sequence models recover the majority of predictable articulatory motion (correlation ≈0.81) relative to acoustic-feature models (≈0.85), and that integrating phonemic information robustly improves speaker-independence and adaptation (Singh et al., 2019, Wang et al., 2023).

4. Mapping Algorithms and Implementation Schemes

The practical transformation from phonetic sequence to articulatory representation can be organized into algorithmic workflows:

Deterministic Vectorization

• CLTS-based Mapping:

1. Parse IPA segment with CLTS to extract descriptive feature bundle.
2. Construct a 39-dimensional vector, applying feature value overwrite rules by specificity hierarchy.
3. For clusters/diphthongs, recursively union or merge feature positives (Rubehn et al., 7 May 2024).

Feature-based Soft Distance Metrics

• Articulatory Edit Distance: Compute Levenshtein distance where substitution costs reflect articulatory feature vector similarity (e.g., weighted Manhattan or cosine distance of features), with type-aware handling for consonants/vowels (Ahmed et al., 2020).

Neural Regressor Architectures

• BLSTM/Attention Mapping:
  • Phoneme input: frame-level one-hot, sometimes with timing expansion or embedding (Singh et al., 2019, Wang et al., 2023).
  • Core model: multilayer bidirectional LSTM or attention-based sequence network, sometimes fused with acoustic stream for hybrid inputs.
  • Target: frame-wise articulatory parameter vector (e.g., EMA 12D, TVs 6D, tongue contour 100D).
  • Loss: MSE, sometimes jointly with phoneme classification (Azzouz et al., 4 Nov 2024, Wang et al., 2023).
• Self-supervised Representation Integration:
  • Use pretrained self-supervised models as feature extractors (e.g., wav2vec 2.0, HuBERT), optionally optimizing on articulatory feature prediction (Tandazo et al., 22 Dec 2025, Lavechin et al., 6 Sep 2025).
  • For cross-lingual generalization, supervision is on articulatory feature vectors extracted from PanPhon/CLTS inventories, collapsing IPA types to feature classes (Tandazo et al., 22 Dec 2025).

5. Applications and Empirical Impact

Phonetic-to-articulatory feature mapping underpins multiple computational and empirical domains:

• Text-to-Articulatory Inversion: Enables direct prediction of vocal tract trajectories or contours from phoneme sequences for speech synthesis, animation, and feedback systems (Azzouz et al., 4 Nov 2024, Wang et al., 2023).
• Cross-Linguistic Comparison: Supplies granular, language-neutral sound similarity metrics for phylogenetic linguistics, loanword detection, and typological studies (Rubehn et al., 7 May 2024, Ahmed et al., 2020, Tandazo et al., 22 Dec 2025).
• Robust and Universal Speech Representation Models: Articulatory supervision in self-supervised models (e.g., MauBERT) yields highly invariant, language-crossing speech embeddings, facilitating adaptation to low-resource or unseen languages (Tandazo et al., 22 Dec 2025).
• Developmental Modeling: Self-supervised neural imitation architectures emulate human-like perceptual-to-motor acquisition by leveraging invariant phonetic representations to learn articulatory control (Lavechin et al., 6 Sep 2025, Siriwardena et al., 2022).
• Automatic Speech Recognition (ASR): Articulatory features support phoneme tying and variant modeling, improving non-native speech recognition, especially in multilingual or accent-robust ASR systems (Wang et al., 2023).

6. Limitations, Extensions, and Theoretical Context

Although current mapping schemes achieve high accuracy, several limitations and research extensions are noted:

• Coarticulation and Motion Dynamics: Linear interpolation of phonological features captures articulatory dynamics up to $r=0.67$ with multi-speaker EMA, indicating that context dependency and coarticulation are critical and benefit from explicit interpolation of context-dependent (unspecified) features (Tandazo et al., 8 Aug 2024).
• Speaker-Independence and Adaptation: Neural architectures leveraging phoneme-only streams yield improved generalization to unseen speakers, but optimal inversion requires joint acoustic-phonemic information (Wang et al., 2023, Singh et al., 2019).
• Feature System Choice and Collapse: Both PanPhon and CLTS systems deliberately collapse narrow phonetic distinctions for scalability, occasionally producing feature-equivalence classes; however, confusion rates are low (e.g., 97.8% of languages in Lexibank have ≤4 non-unique segments) (Rubehn et al., 7 May 2024, Tandazo et al., 22 Dec 2025).
• Developmental Alignment: Computational models confirm the theoretical conjecture that speaker-invariant, phonemically discriminative representations are essential for robust production learning, consistent with infant speech acquisition models (Lavechin et al., 6 Sep 2025).
• Integration with High-Dimensional Articulatory Models: Mapping from phonological to motor commands can be refined via higher-dimensional, physiologically informed synthesis and forward/backward modeling, though further research on full vocal tract state estimation remains active (Siriwardena et al., 2022, Azzouz et al., 4 Nov 2024, Tandazo et al., 8 Aug 2024).

The evolving landscape of phonetic-to-articulatory mapping is characterized by convergence of linguistic feature theory, deep sequence modeling, and application to universal, context-aware speech technology.