AI Pronunciation Feedback for ESL Students

Pronunciation is one of the hardest skills to develop between sessions. A student can review vocabulary from flashcards, practise grammar through exercises, and have a written conversation with an AI partner - but without immediate audio feedback, spoken pronunciation errors go uncorrected for a week at a time.

SpeakRecap addresses this by building pronunciation practice directly into the homework cycle, backed by automatic scoring and a training system that targets each student's specific errors.

How pronunciation scoring works

When a student completes a speaking exercise - either from auto-generated homework or a pronunciation drill - they record themselves using their device's microphone. SpeakRecap sends the audio to a speech analysis model that scores the recording at the phoneme level. The student sees an overall score for each word and specific guidance on sounds that diverged from the target pronunciation.

The model is tuned for English and evaluates against General American pronunciation as the default standard. Scores are percentage-based: anything above 80% is considered good, 50–80% is satisfactory, and below 50% is flagged for extra attention.

Persistent error tracking

A single pronunciation score is useful. A pattern of scores across multiple sessions is far more useful.

SpeakRecap tracks pronunciation errors over time and identifies persistent issues - sounds or words where a student consistently scores low across three or more sessions. These persistent errors are surfaced in your progress tracking view, giving you hard evidence for a conversation that might otherwise feel subjective: "You've been struggling with the /θ/ sound for the past month - let's focus on that today."

Minimal pair training exercises

When SpeakRecap detects a student making the same phoneme error three or more times, it automatically generates a minimal pair listening exercise. Minimal pairs are word pairs that differ by a single sound - such as "ship" and "sheep," or "cat" and "cut" - that are often confused by learners from specific language backgrounds.

The student listens to an audio recording of a word and selects which of two options they heard. The exercise is short, focused, and directly tied to their demonstrated weakness. It appears in the student's practice portal alongside their homework - no additional setup required from you.

Teacher visibility

You can review pronunciation scores for any student across any lesson from your dashboard. The data is there when you want it - particularly useful if a student's speaking confidence has plateaued and you need to understand why, or if you're preparing for a progress review with a corporate client who wants specific metrics.

A note on accuracy

AI pronunciation scoring is a powerful tool, but it works best when used as a signal rather than a verdict. The model is trained on a broad range of speakers and performs well on identifying systematic errors - particularly with consonants like /r/, /l/, /θ/, and /v/ that cause difficulty for learners from specific language backgrounds. For more nuanced accent coaching, the scores provide a useful starting point for the conversation you have in the next lesson.