duckspeak
Say a sentence.
Explore your pronunciation.
Tap to record, then say the sentence
00:00
Use the same sentence. Adds an uncalibrated audio comparison, not a new score. Both files: maximum 2 minutes each and 50 MB combined by default. “Try again” keeps this reference.
Children’s mode: English reference text and up to 2 minutes of audio. Predictions are model estimates.
Listening to your recording
Analysis takes a few seconds on CPU. The very first run also loads the models.
That did not work.
The server could not analyze this recording. Try again in a moment.
Details
Scores are model estimates. A sound without a flagged error may still be mispronounced, and recognition confidence is not calibrated.
Score
Tap a word to see its sounds.
- Expected
- Model heard
Debug evidence, not human-verified errors. Times are approximate greedy CTC frame spans, not forced alignment. Confidence is an uncalibrated peak posterior.
Acoustic traces
Approximate traces placed by index, so timing and phone boundaries are not exact. Energy shows relative loudness and is not calibrated.
Phoneme trace, reference vs model (index-based)
Pitch (F0, Hz)
Energy
Phones are IPA. Recognizer confidence is uncalibrated, not a grade or probability of a learner error.
Historical benchmark aggregate — independent of your recording, not an evaluation of the currently running model.