From Phonemes to Paragraphs: Understanding the Basic English Pronunciation Assessment Question Types

When you say "good morning" in a language-learning app, the screen almost instantly displays a set of scores and highlights the word or phoneme that lowered the result. In the previous article, we broke down this "audio → Multi-dimensional Scores" pipeline. But when developers and product managers begin building an app for the first time, they may get stuck on a seemingly simple question: Which question type should I use?

The question type determines how the reference text is organized and the granularity of the assessment—the engine ultimately decides what can be evaluated.

Pronunciation assessment granularity from phoneme to paragraph

This article explains the basic English question types in one place, helping you understand the best use case for each type and the dimensions each one assesses.

What are basic question types, and how do they differ from advanced types?

The simplest definition is:

There are seven basic English question types:

The key difference from advanced types—Q&A (qa), topic description (topic), story retelling (retell), free speaking (freedom), and others—is that most advanced types have no fixed refText. Their goal is no longer to read according to reference text, but to assess the content of the response itself. If you want to assess a user's ability to read, choose a basic question type.

What is the finest granularity available? Start with the phoneme type (phoneme)

If you want to train users to distinguish /θ/ from /s/ or /l/ from /r/, the key question is whether the phoneme itself was pronounced correctly. This is when you should use the phoneme type (phoneme).

It has the finest granularity of all question types. One or more phonetic symbols serve as the reference text; the user pronounces them, and the engine returns overall scores plus a score for every phoneme.

The scoring information includes:

• overall, accuracy, and integrity, corresponding to the three overall scores: total, accuracy, and integrity
• For each phonetic symbol: phoneme (the target phoneme), hypothesis (the phoneme actually pronounced), score (the phoneme score), and a type label (normal/omitted/repeated)

This is the lowest level in the scoring hierarchy. A word question packages the complete "overall score → word → syllable → phoneme" hierarchy, while the phoneme type isolates the phoneme level for focused assessment. It is especially suitable for:

• Introductory phoneme instruction—mastering individual sounds before memorizing words
• Minimal-pair practice, such as ship / sheep
• Locating the source of a pronunciation problem before correction training

What pronunciation problems can the word type (word) reveal?

The word type (word) is one of the most common entry-level types in speaking apps. The interface usually displays a word or phrase for the user to read, and the scoring engine assesses overall pronunciation accuracy, fluency, integrity, and word- and phoneme-level accuracy.

It is also the source of the example at the beginning of this article. The score for "good morning" runs from the outer overall score to each word, then down to syllables and phonemes:

overall / accuracy
  └─ wordInfo (good / morning)
       └─ syllables (g_uh_d / m_ao_r ...)
            └─ phonemes (g / uh / d ...)

Compared with the available assessment dimensions, the word type already covers four: overall, accuracy, fluency, and integrity, plus word-level stress, which indicates whether the stressed syllable is correct.

In practice, the word type can address several specific needs, depending on the scenario:

• Words, phrases, and pure phonetic symbols are all supported: The reference content can be a regular word or phrase, or an IPA88 phoneme sequence such as /mɑːkɪt/. The same question type can support vocabulary repetition and direct phoneme-level practice.
• British and American pronunciations are both supported: The engine can adapt to either British or American pronunciation. For words such as tomato and schedule, whose pronunciations differ between the two variants, you can specify the variant and score against that standard. You can also choose not to distinguish between them, so either pronunciation is scored against the corresponding standard rather than being penalized for accent.
• Uncommon pronunciations can be annotated manually: If a word has an unusual pronunciation, or if you want to specify how it should be read, you can tell the engine all acceptable pronunciations in advance and have it score against those annotations.
• Designed specifically for children: In addition to the general model, a child-specific speech model is available. It better matches children's vocal characteristics and provides more suitable, learner-friendly scoring.

What does the sentence type (sentence) assess beyond the word type?

Once words are connected into a sentence, the range of pronunciation dimensions expands immediately. The sentence type (sentence) supports up to 90 seconds of audio and is closer to real expression than the word type. It also provides both general and child-specific models.

What it adds is mainly connected-speech dimensions:

• rhythm: stress, tone, sense-group pauses (sense), sentence-final intonation (rearTone, rising/falling), and an overall score
• fluency: pause count (pause), speaking speed (speed), and an overall score
• Word-level liaison: whether linking is expected, whether it actually occurs, and the linking type (unreleased stop, overlap, consonant + vowel, and so on)

Why are these dimensions not assessed in a word question? Because linking, weak forms, pause distribution, and sentence-final intonation only make sense in connected speech. A single word is static; a sentence provides the context required for meaningful fluency and rhythm data.

The sentence type also offers two advanced options that are particularly suitable for coaching products:

• comparison: Pass the ID of a standard recording in the question parameters (customization required) to enable comparison mode, which returns correlation curves for speaking speed, intonation, and volume relative to the standard recording
• realtime: Pass this through the common parameters to enable real-time return, allowing scores to appear while the user is reading for a smoother experience

Why is the paragraph type (chapter) the most comprehensive?

When the text grows into a full paragraph, such as a lesson passage, speech excerpt, or long dialogue, use the paragraph type (chapter). Its audio duration limit is 300 seconds, enough for most scenarios.

Its response structure is also the most layered: overall scoring at the outermost level, followed by sentenceInfo for each sentence and then wordInfo for each word. In other words, it wraps another layer around the sentence hierarchy. Phoneme-level and rhythm scores are disabled by default; enable them manually if you need those dimensions.

It also provides a useful clause_sign parameter. By default, periods, question marks, exclamation marks, and semicolons split the text into sentences. If your text requires different boundaries—for example, splitting only on line breaks—you can specify them yourself.

As paragraphs become longer, omissions, extra words, and repeated reading become more likely. Each wordInfo therefore carries a normal, delete, or repeat label, directly identifying what was omitted or repeated. The paragraph type also supports general and child-specific models and real-time result return.

How do correction types (wordcheck / sentencecheck) differ from the types above?

The previous types focus on scoring. wordcheck (word correction) and sentencecheck (sentence correction) focus on identifying exactly what went wrong. Their parameters look similar to standard types, but the type field becomes a core part of the response:

Specifically:

• wordcheck: Supports only one word. Word-level type can be normal, repeat, or delete. At the finer phoneme level, it also marks insert (an extra phoneme) and replace (a phoneme pronounced incorrectly, such as /θ/ as /s/).
• sentencecheck: At sentence level, word-level type covers normal, insert, delete, and replace. Together with complete rhythm, fluency, integrity, and accuracy results, it provides both scores and precise error locations.

As the names suggest, correction types are designed for pronunciation correction. Coaching and pronunciation-correction apps almost always need these two types. The selection rule is simple:

• If you only need scores → use word / sentence
• If you need to show exactly where the user went wrong → use wordcheck / sentencecheck

What is the new letter-spelling type (wordspell)?

The final type is a new member of the basic matrix: letter spelling (wordspell). It takes a different approach. The app displays the meaning of a word, the user recalls the corresponding English word, and then spells it letter by letter—for example, seeing the prompt “cat” and saying “C-A-T.”

Its response is simpler: it returns the letters actually spoken, their confidence, and the start and end time of each letter. Typical scenarios include online spelling competitions and phonics instruction. Note that it only supports spelling one letter at a time and does not currently support combined forms such as double or triple.

For this type, there is no need to focus on scoring dimensions. Think of it as a specialized entry point for spelling scenarios within the basic question-type matrix.

Why do scoring dimensions differ across question types?

The reason is straightforward: the question type determines the assessment granularity, and the granularity determines which dimensions can be calculated. Consider these four types:

The progression is clear: the broader the unit, the more connected-speech features can be observed. A word has no context, so linking and sentence-final intonation cannot be calculated. In sentences and paragraphs, rhythm and fluency finally have enough data. Rhythm scoring requires special annotation to specify the intended rhythm; a later article will explain this in detail.

What three questions should you ask before choosing a type?

As a product manager, start with three questions:

1. How long is the text, and how long can the recording be?
   • One or two words → word (20 seconds)
   • One or two sentences → sentence (90 seconds)
   • A full lesson passage → chapter (300 seconds)
2. Do you care about overall scoring or correction?
   • Scoring only → standard types (word / sentence / chapter)
   • Locate errors → correction types (wordcheck / sentencecheck)
3. Are the users adults or children?
   • Adults → standard model
   • Children → word_kids / sentence_kids / chapter_kids

If the scenario also involves spelling or minimal phoneme training, consider wordspell and phoneme.

Conclusion

Basic question types are essentially a reference-text-driven read-aloud assessment system whose granularity ranges from phonemes to paragraphs. Once you understand the phoneme → word → sentence → chapter axis, you understand the capability boundaries of DolphinSOE English assessment. Add the correction labels from wordcheck and sentencecheck, and you can turn scoring into teaching.

If you would also like to understand open-ended question types without fixed reference text—including Q&A, topic description, and story retelling—watch for the next article on advanced types.

DolphinSOE Pronunciation Assessment has been running steadily in the Japanese market and currently handles approximately 100,000 calls per day (100k calls/day). Customers include language-learning apps, schools, after-school programs, and other educational institutions, and the service has been continuously validated in real high-concurrency, education-grade scenarios. We provide free technical consultation for API integration and a one-month free trial. For business inquiries, please contact us.