Inspiration
My father-in-law lives with ALS. I've spent this year building communication tools for people like him, and one lesson keeps coming back: families are told to bank their loved one's voice while they still can, and most start too late. By then the voice has already changed.
But the voice is usually already there, in years of birthday videos, saved voicemails and voice memos on the family NAS. Nobody has the time or the heart to dig through them. I wanted the NAS to do the digging.
What it does
Say It Again is a skill for UGREEN AI NAS built around one person, the Speaker: someone living with ALS, Parkinson's, a stroke, aphasia or hearing loss.
- Who is who, first. The family says who the Speaker is and that they agreed, picks the Speaker's people album in UGREEN Photos, names one recording they know is them, and confirms the names of the people, pets and places around them. Names come only from the family. The skill never guesses who is speaking or who is in a photo.
- The message bank. It finds the Speaker's recordings across voice memos, voicemails and home videos, transcribes them, and lets the family confirm the Speaker's voice in each recording by ear. Every phrase is kept word for word with its source file and the second it starts, grouped by everyday need: greetings, love, family names, jokes and sayings, requests and comfort. Signature sayings heard in more than one recording rise to the top. A Still to record list shows what is missing while there is still time, and a drop-in folder takes the new recordings the family makes from it. Longer recordings, like Dad telling the story of the fish that got away, are kept whole as Stories. The same voicemail saved from three phones counts once. When one phrase exists in several recordings, Best Take ranks the takes on what the NAS reports about each file and transcript, offers the top two, and the family picks by ear. A voice album in UGREEN Photos fills itself with new videos of the Speaker, and the next run banks what is new. The NAS has no timer, so each run ends with a short Until next time note and a suggestion to set a weekly reminder on the family's own phone; the next run opens with what is new since last time. Audio phrases can be played back from their first second.
- Name Practice. For stroke and aphasia: numbered practice sets from the family's own photos, one card per confirmed person, pet or place. The family confirms every photo ("is this June?") before it is used, the answer key is kept apart, and a voice cue can play the Speaker saying the name.
- Captions. For hearing loss: subtitles for the videos the family watches together, downloaded for movies and shows, or made from the transcript for home videos and saved next to the video.
Originals are never cut, moved or edited. Every change to files, tags, albums and notes waits for the NAS's own approval step. Everything stays on the family's NAS.
How I built it
Say It Again is a Markdown skill written to UGREEN's Uliya Skill Development Standard with the official uliya-skill-kit 0.7.7. It orchestrates 39 steps across 4 of the NAS's registered experts (files, photos, knowledge and media) and 7 of its CLIs, using only actions in UGREEN's capability catalog: timed transcription with speaker separation, people albums, condition albums and the photo timeline, duplicate scans with SHA-256 proof, media metadata, file copies through the approval workflow, tags, saved notes, subtitle search and download, and the music player's seek. Every step uses only real results from earlier steps, and every one of the five delegation states is handled.
It passes the kit's checker with 0 errors and 0 warnings. An independent review against the kit found 22 issues, all addressed, and a second review of the additions found 9 more, all addressed. A labeled simulation on synthetic test data, with computer voices and drawn photos, ran each part up to its approval step: 26 checks passed, 8 need a real NAS, none failed, and not one file changed.
Challenges
- Who is speaking? A NAS can separate voices but cannot know which one is Dad, and there is no voiceprint matching in the catalog. The answer was to make "who is who" the first step and let the family's ear decide, with a short sample and a recording they know is him.
- No scissors. There is no audio cutting action, so each phrase points at its whole recording and its start second instead of a clip.
- Honest rankings. Best Take can only see what the NAS reports about a file and its transcript, not how a voice sounds. So it ranks on those facts, says so, and leaves the choice to the family's ear.
- Long segments. In the simulation, a transcriber returned a whole memo as one segment, burying "love you, kiddo". Each sentence is now its own phrase that keeps the segment's real times, so nothing is made up.
Accomplishments
- It serves people who are easy to overlook: those losing their speech or hearing.
- It never guesses a name, a voice or a time, and it never touches an original.
- It runs on what a family already owns, in their own home.
What I learned
The most valuable thing a family has is often already saved and simply out of reach. And the most important design rule for a skill like this is to know what only the family can know, and ask them.
What's next
Prove it on a real UGREEN AI NAS on the test day: transcription of home videos, speaker labels, and playing a phrase from the bank. Then make it easy for a speech therapist to bring the bank into a speech device.
