Skip to content

02 — What the ear can and cannot do

This is the section that most directly shapes the sentence-level style of the output.

2.1 The transient information effect — [A] Robust, and the central constraint

Written text is permanent; spoken text is transient. Sweller and colleagues showed that the classic modality effect (audio + picture beats text + picture) reverses once the spoken segment gets long: with longer presentations, audio narration became worse than written text, because the listener must hold everything in working memory with no ability to re-inspect it (Leahy & Sweller 2011; Singh, Marcus & Ayres 2012; Wong et al. 2012).

The recommended remedy in that literature is explicit and simple: present spoken information in small chunks, or don't present it in speech at all.

Singh et al. showed directly that segmenting spoken text removes the disadvantage — the transient information effect is a function of segment length, not of speech per se.

Implications. - A hard cap on the amount of new information per uninterrupted speech run (target ≈ 45–90 seconds of narration per segment, with an explicit boundary marker and a beat of silence). - Long, subordinate-clause-heavy academic sentences must be split. A sentence a reader parses by re-scanning it twice is simply lost in audio. - Anything that would require the listener to "look back" must be replaced by explicit restatement. → Rules SEG-*, SENT-*.

2.2 Working memory capacity

Working memory holds only a handful of novel elements at once (~4 chunks for novel material), and element interactivity — how many pieces must be held simultaneously to understand something — is what actually determines load. A definition with three interacting parts is not three facts; it is one high-interactivity item that must be built up piece by piece.

Implications. Never introduce more than ~2–3 genuinely new terms inside one segment; if the source does, split the segment and add a scaffold. Build multi-part definitions incrementally ("first, ... ; now add to that, ...") rather than stating them whole. → Rules LOAD-*.

2.3 Listening vs. reading comprehension — [B]

The evidence is genuinely mixed and worth stating honestly:

  • Rogowsky, Calhoun & Tallal (2016) found no significant difference between reading, listening, and both, for narrative non-fiction comprehension.
  • Other work finds listening worse, particularly for exact detail recall (names, numbers, precise phrasing) and after a delay, while gist comprehension is roughly equivalent.
  • Attention is a larger factor for listening: listeners are typically doing something else, in a noisier environment, and cannot notice that they have drifted.

Implications. Two design consequences follow. 1. The format is fundamentally fine for gist and structure; it is weak exactly where our source material is richest — precise values, names and symbols. So precise values must be selected, reduced and repeated rather than transmitted faithfully (see 05). 2. Attention lapses are the norm, so the document must be re-entrant: frequent orientation cues ("Section 3 of 7. We are still on the calibration problem.") let a drifted listener rejoin without rewinding. → Rules ORI-*.

2.4 Speech rate — [B]

Conversational speech is ~140–180 wpm; ~180–200 wpm is often cited as comfortable for a native listener. Comprehension holds up to roughly 270 wpm and then falls off steeply (~315 wpm). Efficient listening rate tracks an individual's reading rate closely (Rubin et al. 2021).

Implications. mimem should not fight the TTS engine's rate; instead it should budget words. All duration estimates use a configurable wpm (default 155 for technical narration, lower than conversational because of pauses and dense content). Duration budgets drive segmentation, spacing intervals and the elaboration budget. Difficulty is expressed as more words and more pauses, not as a rate change, because rate changes mid-stream are jarring and unevenly supported by TTS APIs. → Rules DUR-*.

2.5 Pauses and silence

Silence is not dead time in audio learning; it is where retrieval and imagery happen. A retrieval prompt without a pause is just a rhetorical question. Practical constraint: current commercial TTS engines vary in SSML support — ElevenLabs and similar neural engines honour break tags but ignore most other SSML — so the only prosodic control we should rely on is explicit break markers plus carefully chosen punctuation and sentence length.

Implications. Pauses are first-class objects in our script model, with a duration, and are rendered either as SSML breaks or as engine-specific markers. Never rely on <prosody>, <emphasis> or <say-as> surviving the engine — normalize everything in text instead. → Rules PAU-*, TTS-*.