Evaluating Novel Speech Input for Smart Glasses
De-risked a new technology investment by establishing that the input is physically feasible and serves unique use cases.
An investment in a new speech input was moving ahead on internal viewpoints, without user evidence
The organization was envisioning a novel speech input for smart glasses. There was no primary data on whether people could physically produce it, or how it differed from the speech types users already rely on.
My ownership
Set the study goals and aligned research directors, PMs, SWEs and linguists on them
Designed the study and ran AI-enabled analysis
Delivered the report and roadmap next steps
To keep investing, two things had to be true: people can physically produce it, and they’d choose it over other voice inputs
I framed the study around those two conditions. Feasibility: can users produce novel speech with sufficient comfort and physical ease? Desirability: what use cases does it serve that existing voice input doesn't? Five research questions broke them down.
Physical feasibility
How easy or difficult is it to physically mouth novel speech compared to other speech types?
Effort & performance
How effortful is it, and how accurate does pronunciation feel?
Learning curve
Is there a learning curve? Does it change over time?
Value driver
What scenarios would this be helpful in?
Social acceptability
How socially acceptable is it, and how is discomfort mitigated?
Task-based lab study with surveys to measure effort and ease; interviews to explain it
Effort and ease had to be measured, not just described, so a structured task with standard surveys measured feasibility against other speech types. Value and social acceptability needed the why, so interviews followed the task.
Participant profile
- 12 internal employees
- Voice UI users
- Early technology adopters
- No knowledge of the workstream
- Available for an in-person study
- Sized for thematic saturation and statistical trend
Methodology
Analysis
Google Sheets → RStudio sanity check → ggplot plots per speech type
Transcription → codebook → first-cycle coding → affinity mapping → synthesis → report with quotes
Tools: Gemini, NotebookLM, Google Sheets
Interviews ran right after the task; AI tooling for the test-bed and analysis
Built the on-screen test-bed with lovable.ai
Recruited employees with no knowledge of the workstream; flagged social acceptability for external follow-up
Sized for thematic saturation and statistical trend; findings positioned as directional
It's feasible and wanted in specific situations, but effortful at first, with social barriers to solve
It's feasible, but effortful until people get used to it
Novel speech production is physically and experientially distinct. It feels effortful and cognitively demanding at first, then improves with familiarity.
“The first few sentences took real concentration. By the end, I wasn’t thinking about it nearly as much.”
Participant · illustrative quote
- Quantify the difference (dB levels, mouth and facial movements) with research scientists to train AI models.
- Map production experience and compensatory mechanisms to reduce cognitive demand.
People choose it over voice input in specific situations
Unique contextual factors and specific use cases motivate users to choose novel speech over traditional voice inputs.
“There are times I’d never talk out loud to my glasses. That’s exactly when I’d use this.”
Participant · illustrative quote
- Identify top use cases and contextual factors so AI models cater to key needs.
Social discomfort is a barrier, but users named clear ways around it
While social discomfort presents adoption barriers, users identified clear and actionable mitigation techniques.
“It felt a little awkward with people around at first, but I could see ways to make it work.”
Participant · illustrative quote
- Expand social acceptability questions to external participants.
- Map mitigation techniques to different scenarios.
The program moved forward on evidence, and engineering got a direction for model training
- Replaced opinion with evidenceEstablished user feasibility of novel speech production, instead of indexing on subjective points of view.
- Reinforced the investment planReinforced the program's investment plan by establishing desirability through value drivers and unique use cases.
- Guided AI model trainingGave engineering direction for AI model training to improve performance with novel speech.
Next time, I'd flag small-sample limits up front and track learning per repetition
What worked
- Rapid execution for directional insights
- AI tooling to build the test-bed
What I'd do differently
- State the interpretability limits of a small sample up front.
- Visualize SEQ and NASA-TLX per repetition to answer the learning-curve question.
*Illustrative visuals and quotes. Study data under NDA.