Accessibility First: Using Generative Audio to Improve Screen Reader Context in NovaMail

AI · 6 min read

NovaMail, a privacy-focused email client startup, tackled a recurrent accessibility problem: screen reader users struggled to get quick context for long or threaded emails. The team prototyped short audio summaries generated by on-device models that prepend a concise 10-15 second brief before a user navigates a thread.

Design decisions prioritized consent and verifiability. Summaries were opt-in by default for accessibility settings, clearly labeled as AI-generated, and accompanied by a confidence score and the option to hear the raw subject line and first message. The team set an internal accuracy threshold and routed low-confidence cases to a 'read full thread' fallback rather than provide a misleading summary.

In beta with 600 users, the feature reduced time-to-decision on triage tasks by 38% and increased users' ability to prioritize high-value messages. Qualitative interviews highlighted that users appreciated the control mechanisms: toggles to hear a summary again, skipping to important replies, and flagging poor summaries for model retraining.

Challenges included ensuring privacy for hosted accounts and avoiding hallucinations that could misrepresent sender intent. NovaMail mitigated this by running models locally on modern devices and by keeping a transparent feedback loop for users to improve the training set. The case demonstrates that generative AI can enhance accessibility when paired with clear UI affordances and conservative accuracy guardrails.