Skip to main content
Sight interpretation — reading a source sentence and interpreting it out loud, immediately — is the core drill of interpreter training. InterpScout’s practice loop is built around one design choice: the model that grades a student never sees a transcript prepared by someone else. It listens to the actual recording.

​
Where the sentences come from

Practice sentences aren’t written by hand. When a course is produced, its aligned paragraphs are split into clean, complete sentence pairs and stored alongside the course. The practice station pulls from that pool — always a full sentence, always tied back to a real course a teacher chose to produce, never a synthetic drill sentence.

​
One round, end to end

1

A sentence appears

The student sees a source sentence (Chinese→English or English→Chinese, depending on the practice direction) and its position in the set (“sentence 12 of 34”).
2

The student records

They speak their interpretation into the browser microphone, listen back if they want, and submit. The recording — not a transcript — is what gets sent for scoring.
3

A multimodal model scores the audio directly

A multimodal model receives the audio clip along with the source sentence, a reference translation, and the course’s official terms for that sentence, and scores it on three dimensions: accuracy (information coverage — what was said, what was dropped, what was mistranslated), term (whether the official renderings were used), and fluency (how natural the delivery sounded). It also writes out what it heard, in text, and an item-by-item diagnosis — not just a number.
4

Missed terms surface immediately

The transcript the model produced is checked, term by term, against the course’s official glossary for that sentence — a deterministic keyword match, not a second model judgment call. Every official term the student didn’t use ends up on the missed list shown with the score.
5

Missed terms become flashcards

A missed official term becomes a candidate for that student’s personal flashcard queue. The next time they open Flashcards, new cards are pulled from the concepts they’ve most often missed recently — the ones already scheduled for review aren’t duplicated.
Language is checked first, mechanically framed into the prompt: if the practice direction is Chinese→English and the student’s audio is in Chinese, the attempt scores zero with an explicit note, before any of the three dimensions are evaluated. This catches the most common practice mistake — answering in the source language — without relying on the model to notice it unprompted.

​
Why direct audio, not audio → transcript → text scoring

Running audio through a separate speech-to-text step first and grading the transcript is the obvious architecture, and InterpScout doesn’t use it. A transcription step discards exactly the signal fluency scoring needs — pacing, hesitation, intonation — and it adds a place for errors to compound before scoring even starts. Sending the audio straight to a multimodal model keeps that signal intact and removes a compounding source of error.

​
Spaced repetition, not a static deck

Flashcards use FSRS (Free Spaced Repetition Scheduler), the same algorithm family behind modern spaced-repetition apps: each review — graded Again / Hard / Good / Easy — reschedules that card’s next due date based on the student’s actual recall history, not a fixed interval. New cards are capped per day so the queue doesn’t balloon, and a card already due for review is never silently replaced by a fresher miss.

​
The progress page

/practice/progress reads from the same attempt history as everything else: a score trend over recent sessions, a practice-day streak, and a ranked list of the student’s weakest terms — the same aggregation logic that powers the class-wide weak-term board a teacher sees (see Classroom). One scoring event, consumed twice: once to schedule the next flashcard, once to draw the curve a student sees when they ask “am I actually improving?”