Why Language Models Struggle with Connected-Speech Transcription

Large language models appear to produce IPA primarily by pattern-matching against training data rather than by reliably executing phonological rules. For word-level transcription this often matters little. For connected speech, where sounds change based on phonetic context, their output can be inconsistent and should be checked against a reliable pronunciation source or authentic audio before you practice against it.

Type the sentence "I should have told him about it" into any AI language model and ask for a connected-speech IPA transcription. The model will return something close to /aɪ ʃʊd hæv toʊld hɪm əˈbaʊt ɪt/. Every word is recognizable. The transcription looks authoritative. It is also wrong. More precisely, it is the transcription of a sentence that no native speaker of American English actually produces at normal conversational speed.

A phonetician would write it closer to /aɪ ˈʃədəv toʊld ɪm‿əˈbaʊt ɪt/. "Should" reduces from /ʃʊd/ to /ʃəd/ as its stressed vowel weakens to schwa; "have" loses its initial /h/ and reduces to /əv/, yielding /ˈʃədəv/ for the whole phrase. "Him" loses its initial /h/ in unstressed position, surfacing as /ɪm/. The liaison mark ‿ between /ɪm/ and /əˈbaʊt/ shows where the two words connect without a break in the airstream. Neither change is optional or stylistic. They are the predictable output of phonological rules that apply whenever these words appear in unstressed positions in fluent speech. The AI model's transcription is not wrong the way a typo is wrong. It is wrong the way a dictionary is wrong when you ask it to describe a conversation.

This distinction between citation-form transcription and connected-speech transcription sits at the center of a much larger question about what language models can and cannot do with phonological knowledge. The answer matters not just for pronunciation learners, but for anyone thinking carefully about where statistical approximation ends and rule application begins. (If you are just looking for prompts that get the best possible IPA out of an AI assistant, that practical guide is here: ChatGPT IPA transcription prompts. This article explains what those prompts can and cannot fix.)

All IPA examples in this article use American English pronunciation unless otherwise noted.

The Two Things "IPA Transcription" Can Mean

Before examining what AI language models do, it is worth separating two tasks that the phrase "IPA transcription" conflates.

Citation-form transcription records how a word sounds when spoken in isolation, as you would find it in a dictionary. The International Phonetic Alphabet, developed by the International Phonetic Association from 1886 onward, was originally designed for exactly this purpose: a stable, one-symbol-one-sound system that could represent any human language sound without ambiguity. When a dictionary lists "water" as /ˈwɑːtər/ (American English) or /ˈwɔːtə/ (British English), it is giving you the citation form: the word as it would sound if you said it alone, slowly, and clearly.

Connected-speech transcription records how words sound in the context of fluent natural speech. This is governed by a set of systematic processes, collectively called connected speech in linguistics, that alter, reduce, and merge sounds depending on what precedes and follows them. Weak forms, assimilation, elision, liaison, and the American English flap T are not stylistic variations. They are recurring patterns of fluent speech that native speakers produce largely automatically, varying with accent, emphasis, and speaking rate.

The two transcription types serve different purposes. Citation-form IPA is the right tool for learning vocabulary pronunciation and distinguishing minimal pairs. Connected-speech IPA is the right tool for understanding why fluent English sounds the way it does, and for training yourself to produce it.

What a Language Model Actually Does When It Transcribes

Large language models are trained on text. They learn statistical relationships between tokens (words, subwords, punctuation) across enormous corpora. When a model produces an IPA transcription, it is drawing on patterns it has seen in that training data: phonetics textbooks, linguistics papers, ESL learning resources, dictionary entries, language learning forums.

For citation-form transcription, this works reasonably well. The training data contains large numbers of word-level IPA entries, and a model that has seen "water → /ˈwɑːtər/" enough times, in enough contexts, will reproduce it reliably. The same holds for most common English vocabulary. The model is not "looking up" the transcription. It is producing the most statistically likely output given the input, and for common words, that output happens to be correct.

The problem emerges at the sentence level. Applying connected-speech rules requires computing the phonetic environment of each segment: what sound precedes it, what follows it, whether the syllable is stressed, where the phrase boundary falls. Then the appropriate rule from the phonological grammar has to be applied. A language model can approximate this for the most frequent, well-documented cases (the function word "a" reducing to /ə/ appears throughout training data). It produces inconsistent results on less frequent patterns, longer phrases, and combinations of rules that interact at the same word boundary. Crucially, this inconsistency is difficult to predict in advance. The same model may handle one sentence correctly and fail on a structurally identical sentence with different vocabulary.

The errors are not random. They cluster in predictable places: function words (which have weak forms that differ significantly from their citation forms), word boundaries (where assimilation and liaison apply), and unstressed syllables (where elision removes sounds entirely). These are precisely the positions that matter most for sounding natural in connected speech.

The Phonological Rules Where Output Becomes Unpredictable

The connected-speech processes where language models produce inconsistent results are the standard inventory described in any introductory phonetics textbook: Roach's English Phonetics and Phonology, Wells's Longman Pronunciation Dictionary, Cruttenden's revision of Gimson's Pronunciation of English. The rules themselves are well-established. The challenge is applying them compositionally to novel input in a way that is consistent and verifiable.

Weak forms are the most pervasive. English has a large set of function words (prepositions, conjunctions, auxiliaries, articles, pronouns) with both a strong citation form and a weak reduced form. "Have" is /hæv/ in isolation and /əv/ in "should have." "For" is /fɔːr/ in isolation and /fər/ in "waiting‿for‿a bus." "Him" is /hɪm/ in isolation and /ɪm/ after a consonant in running speech. A language model that has seen "him" transcribed as /hɪm/ throughout its training data will produce /hɪm/ by default, because that is the most frequent form in the data rather than an assessment of the phonetic context of the specific sentence.

Assimilation changes a sound to become more like an adjacent sound. The phrase "ten boys" is typically produced as /tem bɔɪz/: the /n/ assimilates to /m/ in anticipation of the following bilabial /b/. "In case" becomes /ɪŋ keɪs/, with the /n/ assimilating to /ŋ/ before the velar /k/. These changes are systematic and predictable from the phonetic environment, but they require knowing what the adjacent sounds are and applying the assimilation rule, not recalling a stored form.

Elision deletes sounds in specific phonetic contexts. The /t/ in "last night" is typically absent in fluent speech: /læs naɪt/. The /d/ in "old man" reduces: /oʊl mæn/. Elision is conditioned by the surrounding consonant cluster and the speech rate. That is contextual computation, and statistical approximation handles it inconsistently.

Flap T, specific to American English, converts /t/ and /d/ to the flap /ɾ/ between vowels when the following vowel is unstressed. "Water" becomes /ˈwɑːɾər/ and "better" becomes /ˈbɛɾər/ in fluent speech. The cross-word environment is less consistent: "get a" produces /ˈɡɛɾə/, with the /t/ flapping across the word boundary onto the following unstressed vowel. This cross-word flap is well-known but underrepresented in training data, and model output here is noticeably less reliable than for within-word cases.

Why Better Prompting Narrows but Does Not Close the Gap

A natural response to these limitations is to ask whether more detailed prompting can close the gap. The answer is: partially, for common cases; inconsistently, for the underlying problem.

You can instruct an AI language model to apply weak forms, to use flap T, to reduce function words. For a short, simple sentence, this sometimes produces a usable result; the model has seen enough examples of these rules applied in its training data to approximate them when explicitly prompted. The improvement is real but uneven. As sentence length increases, as less-frequent function words appear, as multiple rules interact at the same word boundary, the output quality becomes difficult to predict. The same prompt that produces a correct transcription for one sentence may produce a partially correct transcription for a structurally similar sentence with different vocabulary.

This unpredictability is the practical problem for pronunciation learners. A transcription that looks authoritative but is wrong in 2–3 positions per sentence, concentrated in the positions involving weak forms and word boundaries, will reinforce the wrong targets if practiced against repeatedly. The error is not visible in the output. It requires independent verification against a system that applies the rules explicitly rather than approximating them from examples.

This is not an argument that language models cannot improve at this task. It is an argument that the current output is inconsistent enough, in the specific positions that matter most for connected speech, that it should not be used as primary practice material without verification.

What Consistent Connected-Speech Transcription Actually Requires

The alternative to statistical approximation is explicit rule application: a system that takes a phonemic representation of each word, identifies the phonetic environment of each segment in context, and applies the appropriate connected-speech rules in sequence. This is how trained phoneticians work. It is what produces transcriptions that are consistent and verifiable rather than plausible-looking but unpredictable.

The difference in output is not subtle. For a sentence like "I want to go out and get a coffee", a full fluent-speech transcription looks like /aɪ ˈwɑːnə‿ɡoʊ ˈaʊt‿ən ˈɡɛɾə ˈkɑːfi/: words connecting across boundaries, weak forms applied to "to," "and," and "a," and the flap appearing across the "get a" boundary. A rule-based engine applies its covered subset of these processes (weak forms, linking, contractions) deterministically. The same input produces the same output every time, though it does not attempt finer detail like the flap /ɾ/ or the glottal stop. An AI language model, even with detailed prompting, will typically preserve the citation forms of some or all of these function words, and may miss the cross-word reduction environments entirely. The gap is not one of effort or instruction. It is one of consistency.

For pronunciation learners, this gap has a practical consequence. Practicing against citation-form IPA for connected speech does not teach you how English sounds. It teaches you how English would sound if every word were read from a list. The disconnect between that practice model and actual spoken English is one of the primary reasons learners who study IPA diligently still find natural native speech difficult to follow. Practicing with audio alongside the transcription helps close that gap. The IPA they learned was accurate. It was accurate for the wrong thing.

A Note on What Language Models Do Well

None of this is an argument against using AI language models for pronunciation learning. They are genuinely useful for explaining phonetic concepts, answering questions about why a specific sound change occurs, generating vocabulary lists with word-level IPA, and providing metalinguistic commentary on pronunciation patterns. For citation-form transcription of common vocabulary, current models are fast and accurate enough to be a practical first-pass tool.

The limitation is specific: connected-speech transcription at the sentence level, where the output is inconsistent in the positions that matter most: function words, word boundaries, unstressed syllables. Knowing where that line falls is what determines whether a learner uses a language model as a useful supplement or as a primary practice resource. The second use case is where the inconsistency becomes a learning liability, not because every output is wrong, but because the errors are invisible and unevenly distributed.

Understanding the gap between what a language model approximates and what explicit phonological rule application produces is, in the end, the same skill as understanding the difference between citation-form and connected-speech IPA: both require knowing not just what the output looks like, but what process produced it, and whether that process is consistent enough to trust.