POS Tagging for Heteronym Disambiguation in TTS

Heteronyms present a major challenge in Text-to-Speech (TTS) systems because identically spelled words require distinct pronunciations based on context. Part-of-speech (POS) tagging functions as a critical disambiguation layer within the TTS front-end pipeline, assigning grammatical categories to words so that the system selects the correct phonetic transcription. By analyzing the syntactic structure of an input sentence, POS tagging bridges the gap between raw text normalization and accurate grapheme-to-phoneme conversion, ensuring the downstream acoustic model produces natural, intelligible speech.

The Heteronym Problem in Speech Synthesis

In written language, heteronyms share the exact same spelling (graphemes) but diverge in sound (phonemes) and meaning. Examples include:

• "Record": Pronounced /ˈrɛkərd/ as a noun, but /rɪˈkɔːrd/ as a verb.
• "Lead": Pronounced /lɛd/ as a noun (the metal), but /liːd/ as a verb (to guide).
• "Close": Pronounced /kloʊs/ as an adjective, but /kloʊz/ as a verb.

A human reader infers the correct pronunciation effortlessly by interpreting sentence structure and semantics. A TTS pipeline, however, must convert text into phonemes deterministically. If the system relies purely on an unconditioned pronunciation dictionary, it cannot determine which phonetic representation to select.

Where POS Tagging Fits in the TTS Pipeline

A standard TTS pipeline consists of three core stages:

1. Text Front-End: Ingests raw text, performs text normalization (expanding abbreviations and numbers), tokenization, POS tagging, and Grapheme-to-Phoneme (G2P) conversion.
2. Acoustic Model: Takes the phonetic sequence and linguistic features from the front-end to generate intermediate acoustic representations, such as mel-spectrograms.
3. Vocoder: Converts the acoustic representations into an audible waveform.

POS tagging operates immediately after tokenization and text normalization. Before the G2P engine consults a lexicon or runs a neural conversion model, the POS tagger evaluates the sentence to label each token with a syntactic tag (such as noun, verb, adjective, or adverb).

How POS Tagging Resolves Heteronyms

Most heteronyms in languages like English divide along grammatical boundaries. For example, alternating stress patterns frequently distinguish nouns from verbs (e.g., object, present, permit).

When the POS tagger processes a sentence, it applies syntactic context to assign tags:

• "She will read [VERB] the book." \(\rightarrow\) G2P yields /riːd/
• "She has read [VERB, past participle] the book." \(\rightarrow\) G2P yields /rɛd/
• "The tear [NOUN] ran down his cheek." \(\rightarrow\) G2P yields /tɪər/
• "Do not tear [VERB] the paper." \(\rightarrow\) G2P yields /tɛər/

The lexicon used by the G2P engine is indexed not just by spelling, but by word-tag pairs (e.g., lead_NN vs. lead_VB). Once the POS tag is resolved, the lookup returns an unambiguous phoneme string.

Modern TTS systems utilize statistical models (like Hidden Markov Models) or neural sequence-to-sequence models (like BiLSTMs or Transformer-based architectures like BERT) for tagging. These models evaluate surrounding words and dependency structures to calculate the most probable part of speech for ambiguous tokens.

Impact on Downstream Acoustic Synthesis

Accurate POS tagging directly dictates speech quality:

• Segmental Accuracy: Pronouncing the wrong phoneme (saying /wʊnd/ instead of /waʊnd/ for "wound") results in immediate misinterpretation.
• Prosodic and Suprasegmental Realization: Heteronyms often change lexical stress based on their part of speech. Incorrect tagging leads to misplaced emphasis, causing unnatural rhythm and cadence even if the basic vowels remain close.

Limitations of Pure POS Tagging

While POS tagging resolves the majority of heteronyms, it fails when two pronunciations share the exact same part of speech. For instance:

• "Bass": Can refer to a fish (/bæs/) or a low musical pitch (/beɪs/), both of which are nouns.
• "Bow": Can refer to a weapon for shooting arrows (/boʊ/) or the front of a ship (/baʊ/), both functioning as nouns.

In these instances, basic syntactic categorization is insufficient. Advanced TTS pipelines augment POS tagging with word-sense disambiguation (WSD) and deep contextual semantic embeddings to resolve intra-category phonetic differences. Nevertheless, POS tagging remains the foundational, computationally efficient first line of defense for linguistic disambiguation in speech generation.