OpenAI Build Week · Education · 2nd place
Dấu
Built by Robert Huynh
A Vietnamese tone lab with pitch extraction and reference-contour comparison in the shipped browser code. Its server-side coaching was not exercised, and accuracy figures remain the builder’s measurements.
I also participated in this hackathon with Trajectory. This coverage was prepared independently after results were announced.
The idea worth stealing
Let the math grade and the model teach.
Vietnamese speech recognition tends to hear the word you meant, not the tone you produced; in the builder’s own test it turned a spoken mả into mã. An app graded by a language model would tell you that you were right when you were wrong. So Dấu grades with a pitch contour and a warping distance, then hands the verdict to a coach, Cô Dấu, to explain what your voice did and how to move it. The model never decides what you said.
What we checked
Date checked: .
Devpost submission and shipped browser bundle and worker; server-side coaching and transcription were not exercised.
Labels apply to each finding and the evidence described, not to the project as a whole.
Not independently verified means the available evidence was insufficient to confirm a claim. It does not mean the claim is false.
-
The tone judge is real signal processing, and it runs in your browser
Verified
The shipped pitch worker resamples to 11,025 Hz, runs a YIN estimator with a 0.22 threshold and parabolic refinement, then scores the contour with a constrained dynamic-time-warping pass against the reference. No audio leaves the browser to be graded. Source
-
The two failing references really were withheld
Verified
The bundle ships exactly 36 reference targets, 19 Northern and 17 Southern. The submission says 36 of 38 passed validation and the rest were held back; the count matches. Source
-
It would rather ask you to record again
Verified
The worker refuses to grade audio that is too quiet, clipped, too short to hold a vowel, or short on voiced pitch, each with a plain-language message. The pitch estimator itself returns nothing when its confidence drops below 0.3 instead of guessing. Source
Under the hood
- Ten numbers describe a tone
- Every contour is resampled to 64 points and reduced to start, end, slope, curvature, range, minimum, dip position, recovery, final rise, and duration, each weighted. An energy signature (a central dip, voicing gaps, a terminal drop) helps separate the creaky ngã and nặng tones.
- Octave slips are fixed before judging
- Pitch trackers jump octaves. The worker converts frames to semitones, compares each against a local median, and snaps outliers a full twelve semitones off back into place before any feature is measured.
Nuance
No repository is linked despite the MIT license, so we read the shipped bundle and worker, not source. The coach, drill generator, and dialogue transcription live behind server routes we did not call; the bundle names gpt-5.6-sol and gpt-4o-transcribe. The 91.7% held-out accuracy figure is the builder’s own measurement.
Sources
Last updated: . Editorial updates do not imply a new technical check.
Builder credits reviewed: , using the official announcement, submission, and any linked credit sources.
Builder? Add context, request a correction, or ask for a re-check →