BabelVoice: Evidence-First Voice Workflows

Inspiration

Voice tools are fast, but their output is usually a black box. In technical, legal, medical, and multilingual work, one changed negation, number, name, or term can change the meaning. BabelVoice starts from a stricter premise: transcription and translation should remain observable, reproducible, and attributable to source evidence.

What it does

The primary contest demo is VoiceGrab, one complete Windows workflow:

• press Right Ctrl in any application to start and stop capture;
• see recording state, route, options, and segment timing in a compact overlay;
• dictate long thoughts while BabelVoice segments audio without ending the session;
• transcribe, review with a GPT-5.6 route, optionally translate, and paste into the active field;
• retain separate recognized, reviewed, translated, and delivered evidence with hashes, model identity, latency, and timeline events.

Quality Lab supports the demo by recording which models are qualified for which product stages. A two-contour Live Interpreter is included as a clearly labeled beta, not as the primary reliability claim.

How we built it

BabelVoice is a TypeScript monorepo with a self-contained Windows .NET audio host. Its local Observer is both the composition root and the operator surface. The workflow is an explicit stage graph: capture, segmentation, ASR, review/domain handling, optional translation, delivery, and evidence storage.

GPT-5.6 performs the post-STT evidence route through BabelVoice's loopback-only managed CLIProxyAPI sidecar. It can review the transcript, classify domain and terminology, prepare final source text, and translate while preserving separate stage artifacts and immutable prompt/model identity. Codex was used for implementation, architecture review, test repair, packaging, and release verification.

What was added during OpenAI Build Week

BabelVoice existed before the event. Judges should evaluate only these material post-kickoff extensions:

• the explicit VoiceGrab stage graph and fused GPT-5.6 review/translation route;
• hardened native overlay, persisted options, hotkey recovery, and delivery evidence;
• focused VoiceGrab acceptance tests and Windows packaging;
• Quality Lab mode/capability evidence;
• Live Interpreter contour and bilingual-timeline extensions;
• judge-facing scope, verification, and privacy receipts.

The immutable baseline and contest source commit are documented in docs/CONTEST_SCOPE.md.

Challenges we ran into

The hardest problem was not recognizing speech. It was preventing hidden fallbacks and ambiguous ownership when one model can cover several stages. We made route ownership explicit and retained separate evidence for every semantic transition, including failures, cancellation, recovery, and final delivery.

Accomplishments that we're proud of

• One hotkey completes the primary workflow across Windows applications.
• Long capture is segmented without ending the operator session.
• Final text is pasted and retained in the clipboard.
• GPT-5.6 participation is visible in route evidence rather than hidden behind a generic provider label.
• Existing product work and contest-period work are disclosed separately.
• The private judge build has repeatable typecheck, build, unit, VoiceGrab, packaging, vulnerability, and secret-scan receipts.

What we learned

A model name is not a product guarantee. A useful voice system needs mode-specific and stage-specific evidence: accepted input, output semantics, latency, prompt contract, delivery proof, and failure classification. The same graph that improves engineering decisions also lets the operator verify the result.

What's next for BabelVoice

• complete human functional and stability acceptance for both Live Interpreter contours;
• expand the human-approved RU/EN benchmark corpus;
• package the virtual audio driver for clean-machine installation;
• add document translation, OCR translation, and local/offline routes without changing the evidence contract.

Built With

• codex
• dotnet
• gpt-5.6
• openai
• typescript
• websockets
• windows