speaker

Speaker is a narrow tool that does one job with unusual rigor: it writes speaker notes that are grounded in what the audience will actually see. The extraction layer is the differentiator. Most note generators read text boxes and miss the rest; Speaker builds a visible-element inventory per slide covering titles, body text, placeholders and text boxes, table row and column text, and native chart titles, categories, series, values, axes and legends, with an OOXML fallback that pulls SmartArt and grouped-shape text python-pptx does not expose. It then renders slides to PNG, runs region-scoped OCR on pictures and media via visual_inventory.py --ocr-scope image-regions with an automatic full-slide fallback, and assembles a compact vision-review packet for a vision-capable agent or a human reviewer. The result is a chain of custody: every spoken sentence traces back to a visible element. Pacing is where v0.8 does the most practical work. The model is deterministic rather than a black box: English lands near 110 wpm, Chinese near 165 characters/min, with time reserved for transitions and [PAUSE] marks. The README is candid about the failure it fixes — a script labeled 15 minutes that ran to roughly 1,800 words and overran toward 30 minutes now targets 1,300–1,400 words. Per-slide budgets and a timing table reporting words/characters, budget and pauses with a TOTAL row make the overrun visible before you are standing at the podium. The glossary toggle, read_slides.py --mode compact, and vision_review.py --format compact keep intermediate artifacts from ballooning, and the default delivery is a summary plus file paths rather than a wall of script in chat (reply show notes to print the full notes). Where it does not fit. This is a Codex skill, invoked from the command line, distributed as speaker-v8.skill with internal skill name ppt-speech-writer. There is no GUI, no cloud service, and no real-time collaboration on notes; documentation is primarily the README plus a Chinese-language README.zh.md. Note generation is driven by internal heuristic pacing models rather than a named external model, so do not buy it expecting per-slide AI reasoning quality to be the selling point — the selling point is evidence coverage and time control. It runs locally from the .pptx with no cloud upload, which matters if your deck is under embargo or contains unpublished results, and it costs nothing under the MIT license. If your decks are plain bullet lists with nothing visual to inventory, you will pay the setup tax for coverage you do not need.