listening-trainer

Listening Trainer

Text-first listening practice with optional zero-dependency TTS. Paths are relative to ${CLAUDE_PLUGIN_ROOT} (if unset, resolve relative to this file).

When to use

The user asks to drill listening, or brings their own audio/transcript and wants exam-style questions on it.

Steps

1. Identify exam + level + part. Load data/exam-formats/<exam-id>.md for the part's shape (speakers, question count, question types). Seed shapes: data/item-bank/seed/listening-scripts.md.

2. Generate an ORIGINAL script (dialogue or monologue per the part) and its questions. Preview timing is exam-conditional: IELTS and the CEFR B1–C2 (Cambridge) exams give time to read the questions before the audio, so show the questions first; TOEFL iBT does not — the questions come after the recording (listen/take notes → then answer), so for TOEFL keep the questions hidden until playback ends. Keep the script hidden until the drill is scored. Match the script to the recording's real duration: a spoken passage runs ~130–160 words per minute, so a ~4-min monologue/interview is ~550–650 words and a short 30–40 s exchange is ~80–110 words. A script that is too short makes the drill easier than the real test. Count the script mechanically before delivering it — write it to a temp file and run wc -w (see the reading skill for the pattern); a self-estimated count is the failure this guards against.

3. Deliver the audio, best available mode. First, set the number of plays from the exam: IELTS and TOEFL iBT play each recording once; the CEFR B1–C2 (Cambridge) exams play each recording twice (their format files say so). Honor that count in every mode below — "once" means once for IELTS/TOEFL and twice for cefr-b1…cefr-c2.

   • TTS (optional, macOS): if the say command exists (command -v say), speak the script at natural speed for the exam's number of plays. First strip the formatting: remove markdown (**, >) and the speaker labels (e.g. > **Receptionist:**) so the synthesizer voices only the spoken words, not the markup. Write the cleaned, label-free plain text to a temp file in a temp location (e.g. tmp="$(mktemp -t listening)"), then play it, e.g. say -f "$tmp" (or say -o "$tmp.aiff" -f "$tmp" && afplay "$tmp.aiff"). For a multi-speaker script, split it by speaker and feed each speaker's turns to a separate say -v <voice> call so the voices alternate. Delete the temp script/audio after playback (rm -f "$tmp" "$tmp.aiff") so nothing is left on disk.
   • Per-item mode (TOEFL Listen and Choose a Response): these are not one script — each item is an independent ~5-second prompt heard once, with the four options read on screen, not aloud. Pre-render every prompt to its own file before the drill starts, then play them one at a time, collecting the answer after each and never replaying:

     d="$(mktemp -d)"; i=1
     # one say -o per prompt, then: afplay "$d/item$i.aiff"
     

     Pre-rendering matters because synthesis pauses would otherwise leak thinking time into a task whose whole difficulty is that it goes by once. Delete the directory (rm -rf "$d") when the drill ends.
   • User-side audio: the user has any TTS or a practice partner → give them the script to synthesize/have read aloud, for the exam's number of plays.
   • Fallback (always works): timed read-through — reveal the script, instruct the user to read it (once for IELTS/TOEFL, twice for the CEFR exams) at natural pace without scrolling back, then hide-and-answer. Say plainly this trains reading-listening hybrid comprehension, not pure listening.