This is a carve-out, here we share topics that may or may not be related to the DARABARA shop, or the items we stock.
===>
Lately I noticed there's a visible uptick in people sharing a speech-to-text work style, mentioning tools like Wisprflow and improved OS-level ASR (Automatic Speech Recognition), especially on Mac OS 27 Beta, and more people expressing their growing habit of talking to AI agents instead of typing to them. I took on the trend and started working with Codex more frequently via voice input while I was on another project. Working through this, a few threads stood out.
I believe it's because multiple things are compounding
First, ASR has genuinely gotten much better, including at something it used to struggle with:multilingual, code-switched speech. Many users, grew up and living in a multi-lingual environment, communicate daily with different languages mid-sentence, such as English-Japanese, Korean-English, and so on, I used to find that most dictation tools choke on that. For example: iOS built-in dictation. But lately that friction has mostly cleared. You can hit the mic and just talk, pulling two minutes of raw thought out of your head, mixing some non-English adjectives or verbs in between the sentences, and the agent can still accurately parses it, strips filler words, and turns it into actionable structure. I am not saying these tools are perfect and as responsive as a human being, because it does often takes a noticeable pause for the voice to text translation to be processed.
Second, the shape of work has changed. In a blended human-AI workflow, it's roughly: human input, then a long stretch of AI work happening in the background. As a human in that loop, you will be spending more time sitting and thinking, so when a thought arrives, typing it out letter by letter breaks the flow in a way that talking doesn't.
Third, more jobs are moving up to a level of abstraction. When you're writing code directly, precision matters at the character level. Now the human's output is closer to a spec or a stance in an argument, which is something people have always found more natural to say out loud than to type from scratch. (Now, the opposite extreme of this take is the reason why I remain confident in these design tools, such as Figma. It's hard for me to regard the voice or text input more effective over human hand when it comes to precision control of pixels.)
And lastly, Lower token costs means cheaper iteration costs, hence lowers the cost of imprecision. When editing later is easy ("actually, forget that, do X instead"), there's less incentive to compose carefully before speaking. Also, worth mentioning is another technical improvement: We have bigger context windows! which mean an agent can actually hold a long rambling brain-dump and triage it into todos and decisions, rather than losing the thread.
One live counter-force worth naming: the friction of typing forces compression and structure that talking doesn't. It's an open question whether heavy voice-to-agent users will develop new self-editing habits (re-listening, having the agent reflect a structured summary back) to recover that. At least I do, I often end my voice input with "Summarize what I said and xxxx."
What this means for future office design
Return-to-office is speeding up, you have to admit it. I think the likely outcome isn't "open office dies," it's that open-plan stops being the default for everything and differentiates into zones by noise level.
There might be a new format of fully opaque private cubicles built to reintroduce the "what are they doing in there" opacity that open-plan was designed to eliminate. Glass pods let someone visibly be "at work" while contributing zero noise to the room. Visual legibility likely matters more than the acoustic engineering, which is largely solved already.
