OpenAI Build Week · Apps for Your Life · 1st place
Second Voice
Built by Ravitez Dondeti
A builder-described communication aid for dysarthric speech: the builder describes candidate sentences and confirmation before speech output. We could not independently confirm the implementation or zero-training claim.
I also participated in this hackathon with Trajectory. This coverage was prepared independently after results were announced.
The idea worth stealing
Personal context is a bias, not a feature.
The builder describes a GPT-5.6 flow that combines a rough transcript, the user’s phrasebook, and the current situation to propose candidate sentences. The builder’s first prompt made recognition worse, and the reason is the lesson: the model let what it knew about the user override what it had actually heard. The redesign puts the acoustic evidence first, structures the output, and tells the model to abstain rather than guess. Then a human picks the sentence, and only then does the app speak.
What we checked
Date checked: .
Devpost submission and public repository availability; no implementation or runtime was independently tested.
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 code is not public
Verified
The submission’s repository returned “not found” during the September 14 check, leaving no public implementation to inspect. The submission also embeds a demo video, linked here; we did not evaluate it. Source
-
The builder published their own negative result
Verified
The submission reports that the first prompt raised word error rate from 22.3% to 32.0%, and that the fix was evidence-first prompting, structured outputs, and an explicit abstention rule. We could not rerun the evaluation, but publishing the number is rarer than it should be. Source
-
Nothing is spoken without confirmation
Not independently verified
The described flow is two or three candidates ranked by confidence, a single low-effort pick or edit, then text-to-speech. That confirm-before-speak step is the product’s whole safety story, and we have only the description. Source
What we need: Repository access at the submitted commit SHA and a runnable build to test confirmation before speech.
Under the hood
- Transcript plus phrasebook plus situation
- The submission describes combining a transcript, a personal phrasebook, and context to return candidates ranked by confidence. The builder describes weighing the phrasebook against the transcript as the hardest calibration for fragmented speech.
- Abstention over a confident guess
- The builder says that when the acoustic evidence does not support a sentence, the pipeline is instructed to say so rather than produce a plausible one. That rule came out of the failed first prompt, not the original design.
Nuance
Our technical findings come from the Devpost text and the September 14 repository-availability check. A demo video link was identified during the September 15 editorial review; its presence does not confirm the implementation, and we did not evaluate it. The zero-enrollment claim and comparison with Voiceitt remain the builder’s account. Public code would allow a fuller re-check.
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 →