Voice Draft Made for ChatGPT

𝐓𝐢𝐭𝐥𝐞 : 𝐓𝐡𝐞 𝐅𝐞𝐚𝐭𝐮𝐫𝐞 𝐈'𝐝 𝐁𝐮𝐢𝐥𝐝 𝐭𝐨 𝐅𝐢𝐱 𝐕𝐨𝐢𝐜𝐞

6 weeks. 31 users surveyed. 6 interviews. 1 clear behavioral insight. Here's the product solution I designed from research to PRD. 👇

𝐓𝐡𝐞 𝐩𝐫𝐨𝐛𝐥𝐞𝐦 (𝐫𝐞𝐜𝐚𝐩):

Students don't use voice on ChatGPT because voice auto-submits the moment you stop speaking. No review. No edit. No control. For a student mid-exam-prep, one wrong transcription and the whole query is gone. That one failure creates long-term avoidance. The fix isn't better speech recognition. The fix is restoring control.

𝟑 𝐬𝐨𝐥𝐮𝐭𝐢𝐨𝐧 𝐝𝐢𝐫𝐞𝐜𝐭𝐢𝐨𝐧𝐬 𝐈 𝐜𝐨𝐧𝐬𝐢𝐝𝐞𝐫𝐞𝐝:

→ Direction 1: Restore Control — editable voice drafts before sending → Direction 2: Reduce Disruption — quieter, less intrusive capture → Direction 3: Structured Thinking — pause/resume for multi-part queries

I chose Direction 1.

It's the only approach that directly eliminates the primary behavioral barrier. Directions 2 and 3 address secondary friction — they can be layered in as the feature matures.

𝐓𝐡𝐞 𝐬𝐨𝐥𝐮𝐭𝐢𝐨𝐧: 𝐕𝐨𝐢𝐜𝐞 𝐃𝐫𝐚𝐟𝐭 𝐌𝐨𝐝𝐞

Voice input redesigned from instant-submit to draft-first.

The new interaction loop: Speak → see transcription in real time → draft held for review → edit or add more → explicitly send

Voice now behaves like typing: controllable, reviewable, and structured.

𝐊𝐞𝐲 𝐟𝐞𝐚𝐭𝐮𝐫𝐞𝐬:

→ Editable voice drafts — transcription becomes editable text before sending → Real-time transcription — builds trust, reduces fear of errors mid-speech → Pause and resume — supports natural breaks without losing the draft → Mixed input — supplement voice with typing in the same draft → Explicit send — voice is never auto-submitted

𝐇𝐨𝐰 𝐈'𝐝 𝐦𝐞𝐚𝐬𝐮𝐫𝐞 𝐬𝐮𝐜𝐜𝐞𝐬𝐬:

Primary metric: % of study-related queries initiated via voice Guardrail: no regression in overall message send rate or session duration

A/B plan: Control → current instant-send voice Treatment → draft-first voice Split → 50/50 on study-segment users Duration → minimum 2 weeks to account for learning curve

𝟑 𝐭𝐡𝐢𝐧𝐠𝐬 𝐭𝐡𝐢𝐬 𝐩𝐫𝐨𝐣𝐞𝐜𝐭 𝐭𝐚𝐮𝐠𝐡𝐭 𝐦𝐞 𝐚𝐛𝐨𝐮𝐭 𝐏𝐌 𝐭𝐡𝐢𝐧𝐤𝐢𝐧𝐠:

→ Low feature adoption is almost never a discoverability problem. Dig into the behavioral mismatch first. → Non-goals matter as much as goals. I explicitly called out that this feature doesn't try to improve STT accuracy or replace typing — knowing what you're not solving keeps scope clean. → Trade-offs are decisions, not gaps. Draft-first adds one extra step to the send flow. I accepted that trade-off because 77% of users value accuracy over speed — and this step is the entire value delivery.

This is my first end-to-end PM project — market landscape → user research → wireframes → full PRD with A/B plan, metrics, open questions, and trade-off documentation. NextLeap