Go Deep on a Capability Module, Not an Industry: Building a Communication Tool for the Deaf in Ten-Odd Minutes

“Tech for good” is easy to say. Harder to actually do—especially in a way that’s sustainable rather than performative. I worked at Tencent for many years and still appreciate the culture there: high autonomy for employees, and genuine resources poured into public-interest projects. Recently I started a pro-bono project called Chuansheng (传声) to help deaf and speech-impaired people communicate face-to-face with hearing people.

It started during an offline class in Hangzhou. One of the students was sharp enough to notice a problem hiding in plain sight: when a deaf person needs to communicate face-to-face with a hearing person, existing tools barely help. Speech recognition accuracy is one issue, but the bigger one is that even if the other person’s words get transcribed, a deaf person can’t type fast enough to keep up with the pace of a real conversation. The whole interaction flow breaks down.

The solution we came up with works like this: one side does precise speech recognition—turning what the hearing person says into text the deaf person can read in real time. The other side analyzes the conversation context and auto-generates several possible replies. Tap one and it speaks aloud via TTS. The system also stores common phrases—“thank you,” “sorry,” “could you say that again”—each one tap away from being spoken. The whole design lets a deaf person complete a natural face-to-face exchange without typing a single character.

I reused the speech-recognition framework I’d already built for a voice input tool and had the website up in a little over ten minutes, right there in class. It’s live now, being tested on a small scale. I’m covering all the compute costs myself for the time being and plan to reach out to public-interest foundations to make it sustainable long-term.

That ten-minute build made me rethink an old assumption in enterprise tech services: the conventional wisdom says you should go deep in one vertical industry so your frameworks, products, and delivery processes can all be reused efficiently. That logic holds. But my interests are too broad and I’ve taken on projects across many industries—so I ended up on a different path. Instead of going vertical in an industry, I went vertical in capability modules.

Across all those cross-industry projects, I built deep expertise in specific technical modules—image generation and audio recognition being the two clearest examples. These modules don’t care what industry they’re deployed in.

Take audio recognition alone: publishing uses it to transcribe interviews, livestreaming uses it to create clips, short-video uses it to add accurate subtitles, and this pro-bono project used it to ship a high-quality MVP in minutes. One module, four completely different industries, all working.

The conclusion is straightforward: you don’t need to concentrate your research in one industry. Get one capability module above 90-percent quality and you can reuse it across industries and projects on the spot. Going vertical in an industry is one valid path; going vertical in a capability module is another—and for people with broad interests and diverse project portfolios, it may be the more realistic one.

Chuansheng is living proof of this for me. It wasn’t built from scratch as a charity product. It was an already-polished technical module plugged into a new scenario in ten-odd minutes—and it actually helps people. Tech for good doesn’t require starting over. It can be the natural extension of capability you’ve already been building every day.

Source: EP0084 (2026-08-24, 2:44) · audio recorded by the host · on-screen presenter is an AI avatar · both text and timecodes come from gemini-2.5-pro ASR of that recording, with three passages re-checked by ear

[00:00] “Tech for good” is never just a slogan When it comes to AI for public good a lot of the time, large companies just do it as a side effect I worked at Tencent for many years and I still really like its corporate culture First, employees get a lot of freedom which is how I did so much learning and exploring Second, Tencent really does put a lot of resources into public-good work Recently, I also started a charity project to help the deaf and mute community It started when I was teaching an offline class in Hangzhou

[00:26] One student very sharply spotted a real need Communication between deaf and mute people and hearing people is extremely inconvenient Existing speech recognition tools have fairly low accuracy And even when it does recognise what the other person said a deaf and mute person still can’t type a reply quickly So at the time, we came up with an idea We’d reuse the technical framework I had developed for my speech input method to accurately recognise what a hearing person says

[00:52] show it to the deaf-mute user, and at the same time judge what was said and auto-generate several different possible replies That way, the user can just tap one of the options to reply quickly You can even use TTS to have it spoken aloud The system also stores a lot of common conversational phrases Deaf and mute people just need to tap to have it speak and reply with voice For example, “thank you” or “sorry”

[01:18] “Could you say that again,” things like that some simple phrases This makes it easier for deaf and mute people to communicate with hearing people So during the class, I spent about ten-odd minutes and built this website This got me thinking about something We used to say that when you’re doing this kind of enterprise-service technical development it’s best to go vertical on one industry, right? so that the framework and technology that I develop can be reused, even at the product level But because my personal

[01:43] interests are just too broad plus I like doing different things I’ve taken on many cross-industry projects And I found that in these project development processes I would, on certain vertical technical modules accumulate very deep expertise For example, image generation, right? or for example, audio recognition These are functions that can be reused quickly across industries Just take audio recognition for example The publishing industry can use it to collect and record interview transcripts

[02:08] Live-streaming can use it to cut clips Short videos can use it to add accurate subtitles And even for this charity project it could be used to very quickly ship a high-quality MVP for MVP development and launch you don’t actually have to focus your research on one industry As long as for one capability module you get it above ninety out of a hundred you can reuse it quickly across industries and across projects This charity project is actually live now It’s in small-scale trial use

[02:33] As for all the compute cost I’m paying for it myself for now Later, I’ll also try to approach charity foundations to make it something long-term and sustainable something that can truly help our deaf and mute friends a genuinely useful helper