Designing Onboarding for Atypical Speech

01 — Background / About the App

What Project Relate does

Project Relate is a Google Research app that uses machine learning to help people with non-standard or atypical speech communicate more easily and independently, while giving them access to Google Assistant. Users train a personalized speech recognition model by recording a set of phrases; once trained, the app can transcribe their speech, read it aloud in a synthesized voice, and connect them to voice-activated tasks.

PROJECT_RELATE // GOOGLE RESEARCH MODEL ONLINE

Speech, decoded

Machine learning that lets people with atypical speech be understood and stay independent.

INPUT — □ ×

Non-standard speech

Hard for machines and people

PERSONAL MODEL — □ ×

Trained on 500 phrases

Recorded by the user · on device

OUTPUT — □ ×

Understood by anyone

Text · synthesized voice · assistant

01 / Listen

Transcribes speech in real time

02 / Repeat

Reads it aloud in a synthesized voice

03 / Assist

Runs voice tasks via Google Assistant

04 / Goal

Communicate more independently

PRODUCT_BRIEF · SIGNAL_CHAIN [ RELATE ]

02 — Problem

A funnel that collapsed after signup

Funnel data revealed steep drop-off after signup: only about 20% of invited users ever opened the app, and just 3% became 30-day active users. The steepest declines happened right after users started the app and were asked to record phrases, and again once they had recorded enough samples to build a personalized model. Many users simply didn’t understand what they would get in return for the effort.

PROJECT_RELATE // USER_FEEDBACK.LOG SEV: HIGH · 04 ENTRIES

Problems

> unresolved · awaiting triage

ERR_01 / CLARITY — □ ×

“It’s a bit unclear what I get after all this work.”

ERR_02 / CONTROLS — □ ×

“The start and stop button is a bit confusing. I thought I could just keep talking.”

Ignore Solve

ERR_03 / EXPECTATIONS — □ ×

“If I knew I had to record 500 to get the feature, I would have done it in one sitting.”

ERR_04 / ONBOARDING — □ ×

“There were no instructions how to use the app after signing up.”

USABILITY_REVIEW_V1 · 2026 [ IGNORE ] [ SOLVE ]

PROJECT_RELATE // FUNNEL_TRACE.LOG LEAK DETECTED · 02 STAGES

Drop-off

> where users vanish

LEAK_01 / ACTIVATION — □ ×

Started the app → recording something

73% continue 27% lost

▾27%

Decrease in users

LEAK_02 / PERSONALIZATION — □ ×

Recorded something → enough for a personalized model

84% continue 16% lost

▾16%

Decrease in users

FUNNEL_TRACE_V1 · 2026 [ IGNORE ] [ SOLVE ]

PROJECT_RELATE // RESEARCH_FUNNEL.TRACE EXIT VELOCITY: 97% · 07 GATES

The Void

> 100 enter the funnel · 3 come out

Completed interest form

GATE_01 · Baseline

100%
Invited to app

GATE_02 · ▾ 21 PTS

79%
Started app

GATE_03 · ▾ 59 PTS · Critical

20%
Recorded something

GATE_04 · ▾ 3 PTS

17%
Recorded enough for personalized model

GATE_05 · ▾ 8 PTS

9%
Used their model

GATE_06 · ▾ 2 PTS

7%
30-day active users

GATE_07 · ▾ 4 PTS

3%

RESEARCH_FUNNEL_V2 · WIREFRAME_TRACE · 2026 [ IGNORE ] [ SOLVE ]

03 — Challenge

Built without UX — value never surfaced

Project Relate had originally been built by engineers without dedicated UX support, so it lacked visual polish and a clear way to communicate its value. Users were asked to record 500 phrases with no explanation of why, leaving them confused, discouraged, and prone to abandoning the app before ever experiencing its benefits.

PROJECT_RELATE // CONSTRAINT.LOG CRITICAL · ACCESS BLOCKED

Constraint

> gate detected before value delivery

WARN_01 / TRAINING QUOTA — □ ×

Users must train 500 phrases before getting access to their speech model.

QUOTA_METER — □ ×

500

Phrases required

Reward unlocks at 100%
Partial credit · none
Model state [ locked ]

PHRASE_BUFFER · 060 / 500 — □ ×

Typical user records ~60 440 never recorded

CONSTRAINT_LOG_V1 · 2026 [ IGNORE ] [ SOLVE ]

04 — Solution

Three pillars for first-run

I designed and shipped a new onboarding sequence built around three pillars: feature education, step-by-step onboarding, and re-teaching. Together, these clarified the app’s purpose up front, guided users through the recording process with contextual tooltips, and re-engaged them with a refresher once their personalized speech model was ready.

01 / 14 · Welcome Autoplay

Feature education — Welcome, How it works, and Getting started — 01 / Feature education Welcome · how it works · getting started

Onboarding recording flow with step-by-step tooltips — 02 / Step-by-step Recording tooltips

Re-education after the speech model is ready — Listen, Repeat, Assistant — 03 / Re-teach Model ready

05 — Impact

Shipped — and highlighted at Google I/O

The designed onboarding was featured in the accessibility portion of Google I/O in May 2024, and internal leaders across UX, engineering, and accessibility praised the work for improving usability and informing Project Relate’s product roadmap.

PROJECT_RELATE // PATCH_NOTES.LOG STATUS: RESOLVED · LEAK_01 PATCHED

Patched

> largest onboarding leak closed

27%

Before

Delta ▾8 Points

19%

After

FIX_01 / ONBOARDING — □ ×

Reduced the largest onboarding drop-off from 27% to 19% — more users successfully record their first sample.

Drop-off before27.0%
Drop-off after19.0%
Delta▾ 8 pts
Relative recovery▲ 30%
Continued (start → record)73 → 81

PATCH_NOTES_V1 · 2026 [ SOLVED ]

Google I/O 2024 — Accessibility

06 — Process

My Design Process

Kickoff, journey mapping, research and iterations. Expand for full story

06a — Project kickoff

Understanding the speech experience

I led a cross-functional brainstorming workshop to align the team on user needs across three impairment groups: cognition, motor, and hearing. Using a “how might we” prompt for each group, the team generated and clustered ideas on an affinity map, giving us a shared direction before any design work began.

Kickoff board — user goals, cognition, and motor needs

Kickoff synthesis — user goals and focus areas

User stories table: As a user, I want to, so I

User stories — jobs to be done at first-run

How might we statements for cognition, motor, and hearing

How might we — cognition, motor, and hearing

06b — User journey

Aligning the team on friction

I mapped the end-to-end journey from download to receiving a personalized speech model, documenting tasks, emotions, and opportunities at each stage. Sentiment dropped sharply during the lengthy recording process and recovered once users received their model — pinpointing exactly where onboarding support was most needed.

PROJECT_RELATE // USER_JOURNEY.MAP 05 STAGES TRACED

User journey

> download → consent → record → wait → use

Action

Download Relate

Auto-enrollment consent

Record 500 phrasesEmotion ▾ low

Complete recording

Receive model

Task list

Play Store · download · Gmail sign-in

Read consent · add a helper

Record phrase cards with no payoff in sight

Finish the last cards · wait for processing

Notification · open Listen / Repeat / Assistant

Feeling

Excitement · curiosity

Tedious · unsure of the value

Confusion · discouraged · unsure

Accomplished · anxious about the wait

Rewarded · excited to be understood

Fix here

Set expectation before install

Teach the product before the ask

Step-by-step onboarding · show progress

Set wait-time up front

Re-teach features when the model returns

JOURNEY_MAP_V1 · 2026 [ RELATE ]

06c — Research

Finding the real friction

A heuristic evaluation flagged key improvement areas. User interviews confirmed users were confused about features and personalization, not just “unmotivated.” Funnel analysis also showed the two biggest drop-offs occurred right after app start and after completing enough recordings to qualify for a personalized model.

PROJECT_RELATE // AUDIT_TRAIL TRACE COMPLETE

Research trace

> heuristic · interview · funnel

01STEP_01

Heuristic evaluation

Audited onboarding against usability principles.

OUT Unclear progress feedback and insufficient training.
02STEP_02

User interviews

Spoke with people who signed up and started recording.

OUT They didn’t understand the value of 500 phrases — and got no post-signup instructions.
03STEP_03

Funnel analysis

Traced the onboarding funnel step by step.

OUT Two major drop-offs: right after app start, and after qualifying for a personalized model.

HEURISTIC EVAL — □ ×

CONCLUSION — □ ×

Users abandon before the model can prove it works.

RESEARCH_METHOD · AUDIT_TRAIL [ RELATE ]

06d — Design iterations

Education, pacing, and refinement

Based on the research, I framed four “how might we” directions that became three design pillars: feature education, onboarding, and re-teach. I iterated on low-fi wireframes with a UX writer and UX expert to tighten copy, tooltip sequencing, and pagination, and to address edge cases from research.

PROJECT_RELATE // TOOLTIP_ITER.LOG 04 PASSES

Tooltip iterations

> wireframe → grey → brand icon → final

01WIREFRAME

Map the taps

Arrows only. Prove the sequence: start, read, stop, review, next — before any UI chrome.

PASS_01 / LO-FI — □ ×
02GREY

Place in the real screen

Dark bubbles in the recording UI. Test where they sit — no brand yet.

PASS_02 / MID-FI — □ ×
03BRAND ICON

Make it a tour

Relate mark, step dots, and Next / Finish so first-run feels guided, not guessed.

PASS_03 / BRAND — □ ×
04FINAL

Ship the teacher

Sage bubbles, tighter copy. Dismissible if you already know the flow, plus a review step so you can check the audio before it counts.

PASS_04 / SHIPPED — □ ×

TOOLTIP_ITER_V1 · 2026 [ RELATE ]

06e — Final design

Worth the wait — then ship

When the personalized model was ready, a re-education moment closed the training-to-payoff gap — with wait-time expectations set upfront. Internal testing produced consistently positive testimonials.

Next steps

Partner with marketing on a GM3 campaign for new and existing users
Run usability testing against the GM2 baseline

Lessons learned

State product value early — purpose before effort
Design for edge cases and cognitive diversity from the start
Keep onboarding step-by-step so users can build a mental model

Non-standard speech

Trained on 500 phrases

Understood by anyone

Heuristic evaluation

User interviews

Funnel analysis

Map the taps

Place in the real screen

Make it a tour

Ship the teacher

Have a problemworth solving?

Have a problem
worth solving?