Voice is Underutilized
The author describes how he uses speech for input and output with AI over 2026, from transcribing design meetings to a daily audio briefing built with text-to-speech. He argues voice is a practical, programmable modality that most people haven't tried.
- Transcribed meetings and ramble-style voice memos give agents rich context; a commute brain dump, transcribed and run overnight with /goal, often comes back solved.
- A daily 10-20 minute audio briefing (news, tasks, RSS summaries, local nature and astronomy events) stays interesting only with enough schedule variety to avoid feeling formulaic.
- Voice mode with Claude, wired to MCP tools for tasks, notes, and flashcards, lets him triage work and quiz himself verbally, which he finds harder to cheat on than Anki.
Benjamin Congdon's argument is that voice is an underused way to work with AI, and he breaks it into input, output, and real-time use. On input, he uses speech to ramble at coding agents, which is easier than typing, and he has wrist sensitivity that makes typing a problem. Local speech-to-text is now good enough for his work: he runs NVIDIA's Parakeet models through an app called FluidVoice, then has Claude Opus clean up transcription errors, and he says the result is nearly indistinguishable from perfect for his purposes. His team transcribes design syncs, and engineers then iterate on designs with agents between meetings instead of writing long docs. He also records stream-of-consciousness voice memos on his commute, transcribes them at home, and sets an agent running overnight with /goal. He says the problem is often solved by morning.
On output, he has built a personal daily audio briefing of 10 to 20 minutes, generated with text-to-speech. It pulls in local, national, and international news, stock conditions, aviation weather for flying days, his tasks and calendar, RSS article summaries, generated research segments drawn from his recent context via Exa, local nature events like bird migrations and salmon runs, concerts, and astronomy such as planet visibility and meteor showers. The hard part, he says, was building enough variety into the segment schedule that it doesn't sound formulaic. A plain "give me a daily briefing" prompt goes stale within days. He switched from GPT-4o mini TTS after its deprecation to Gemini 3.8 TTS, which he calls expressive, cheap, and usable through OpenRouter. He also generates explainer podcasts, which taught him a fair amount about aviation and distributed systems.
For real-time use, he relies on the live voice modes in Claude, ChatGPT, and Gemini, mostly on his commute and while walking around the house to triage tasks and capture admin items. He prefers Claude because, after Claude's voice mode added MCP support, he could connect his personal software for tasks, notes, and flashcards. He reviews flashcards by speaking answers aloud, which he finds harder to fake than flipping through Anki, and he argues that saying answers out loud helps retention. His closing pitch is to try recording voice memos on your commute or talking through tasks with a model, since the overall payoff is plenty of practical, everyday value.