Memoria is a mobile app for keeping family stories. You talk about your life, and it turns what you said into text your family can read and keep.
You open the app and answer a question out loud, or say whatever is on your mind. Memoria turns the recording into clean text with a title. Over time, you can gather several memories into one longer story and share it.
My grandmother always wanted to write down her life. By then it was hard for her to hold a pen. She wrote a little over a page and stopped. She never got to finish it.
My grandfather spent decades writing his memoir, San Xian Ji Yi (Memories of the Third Front). I am now translating it into English.
Later I talked with people about it. Often it was the children who wanted the stories. One person had recorded a parent once and never touched it again. Another had hours of recordings of a grandmother who has since passed away, and had not listened to any of them.
They all wanted to keep the stories. But every step took too much effort. Writing a memoir goes the same way. You buy a notebook, decide to start, and put it down after a few pages.
The brown line is the goal, not a measurement.
Three things stop people. I set out to remove each one.
The first version opened on a deck of topic cards. Pick one, and start talking.
It took me two tries to see why testers got stuck here. Before May, three or four of them said they opened the app and didn't know what to do. Some never noticed there were more cards.
I thought they couldn't find the other cards. So in May I made the card slide aside by itself the first time you open the app, to show another one underneath.
The fix for the wrong problem. The card moves by itself to show there are more.
The bigger problem was still there. In July, a tester told me what she wished the app had. First, a home page sorted by category. Later, two modes called Interview and Diary.
Her two ideas pointed at the same thing. Some people need a topic to get started. Others already know what they want to say. With only topic cards, the second group had to pick a topic first.
So the home page now offers three choices. Speak Freely starts recording right away. Find a Topic opens the cards. Follow-ups asks new questions based on what you have already told. It only appears once there is something to follow up on.
The cost: people who want a topic now tap once more. I have no data yet on how much that costs.
The moment someone stops talking, their recording is saved. Everything else can wait.
What went wrong. I built for recordings of two to five minutes. My mother talked about a trip for 83 minutes. When she pressed stop, the app kept loading, then gave up. The recording was lost.
Why. The app saved only after it had turned all the speech into text. For 83 minutes of speech, that took too long.
What I changed. Saving now comes first. The text, the title and the category come afterwards. A recording can now run up to two hours, so nobody has to stop in the middle of a story.
The cost. The text shows up a little later. Until then, the recording says “Preparing transcription”.
Right after stopping a 3½-minute recording. “Memory saved” comes first. The writing out follows. (2× speed.)
A tester in his 60s pressed Pause when he meant to finish. Then he was stuck, going back and forth between Pause and Resume. The finish button sat in a corner and said "Done".
Now, after a pause, a Finish button appears right under Resume. "Done" became "Finish". In later tests I didn't see anyone get stuck again.
A recording is still hard to read.
One user told me the text was full of "um" and "like". Reading it was a bit embarrassing.
The AI can clean up speech, but it can also change what someone meant. So there are two layers. The first is exactly what was said. Nothing changes it, and it is always one tap away. The second is a copy the AI turns into readable text.
I wrote the rules the AI follows. It may remove filler words, in English and in Chinese (那个, 就是, 嗯), and fix grammar. It may not summarize, leave things out, make things up or translate. People can edit the second layer by hand.
Well, I I drove to the US with my wife last weekend. It was a long weekend. Uh so, on the first day, we drove to Seattle. … Uh but actually, it it was pretty like expensive. Uh but anyways, uh some of the stuff were were pretty good, uh attracted her.
I drove to the US with my wife last weekend. It was a long weekend, so on the first day, we drove to Seattle. … It was actually pretty expensive, but anyways, some of the stuff was pretty good and attracted her.
From a recording I made for this page. Red words are gone. "But anyways" stays, because that is how I talk.
After this shipped in April, the same user said the text was readable now.
One small choice: if you edit the clean copy and press reset, you get the clean copy back, not the word-for-word one. The word-for-word one is still there in layer 1.
Delete two paragraphs, then press reset. The clean copy comes back.