FenwriteApp concept · Bitfuzed · 26 Oct 2026
Fenwrite
A transcriber with a fennec fox in it. Fen lies on his dune looking at you. Press record and his ears go up, he picks up his pen and writes down every word in the notebook on his tail, and he keeps track of who said what.
- Live words, with Speaker 1, 2, 3AssemblyAI streaming labels each turn as it arrives.
- A clean copy and a summary when you stopA second AssemblyAI pass fixes who-said-what; OpenAI writes the gist and to-dos.
- Everything saved, on phone, wrist and home screeniPhone and Android, an Apple Watch recorder, and widgets.
Try it: tap the apricot button on the phone. The conversation is a scripted sample; Fen is the real Higgsfield render, chained clip to clip the way the app will do it.
Morning! Ready when you are.Tap the button and I'll start listening.
Recent
Got it. Tidying up my notes…The live words stay readable while this runs.
- Clean copy from the whole recordingAssemblyAI Universal-3.5 Pro
- Who said whatSpeaker labels, plus name guesses
- Summary and to-dosOpenAI, through our Worker
Saturday market plan
The team set up Saturday's stall: a bigger bake than last time, a new six-bun box, and power for the card reader and fridge.
I think Speaker 1 is Maya, Speaker 2 is Daniel and Speaker 3 is Priya. They said so.
Decided
- 60 sourdough and 40 rye
- Cinnamon buns in boxes of six, $12
- Ask the organisers for power
To-dos
- Send the stall list to the organisers by Thursday Speaker 3
- Make the bun-box sign Speaker 3
- Book the van for 6 am Saturday Speaker 2
The character
Meet Fen
Fennec foxes have the biggest ears for their size of any fox, big enough to hear prey moving under the sand. That is the pitch for a transcriber, so Fen's ears do the acting: relaxed while he waits, straight up and turned forward while someone talks.
The home frame. Every "looking at you" clip starts and ends here. 16:9, so it fills a quarter of a phone screen edge to edge, and crops to a third without losing an ear.
- Ears
- His listening meter. Up and forward while recording, one flicks when a new voice starts.
- Eyes
- On you when he's waiting, on the page while he writes, with a glance up now and then.
- Tail
- His desk. He lies curled round it, the notebook rests on top, and the dark tip anchors the frame.
- Neckerchief
- The app's apricot. It is the one colour that still reads on a watch face or a small widget.
- Pen
- Always in his paw, so no clip has to make it appear or vanish.
How he was found
sky kepta · Plain, morningThe cream morning sky blends into a light UI. He sits off-centre and the pen isn't on the page.
neckerchief keptb · Neckerchief, middayGave him the apricot accent. The blue sky felt colder than the app.
droppedc · Reading glassesCharming, but thin glasses wobble between video frames and vanish at widget size.
body keptd · Rounder, duskThe cutest body; the tail wraps round him like a nest, which makes the best desk.Fen is d's body with b's neckerchief under a's sky. Your stock image (Dreamstime) was used only as a render-style reference, the same way as Tora's. Fen is an original character, not a copy of the stock fox.
His poses
Looking at youHome frame A: idle, wave, the app's home screen
WritingHome frame B: the whole recording
ThinkingSummary on its way
HappyNotes ready, notification art
NappingEmpty library, night, widgetsAnimation
How Fen moves
Same method as Tora: Seedance 2.5 image-to-video clips that start and end on a fixed frame. Fen has two home frames, and every clip begins and ends on one of them, so the app can chain clips without a jump.
A · Looking at you
B · Writing
The gallery loops every clip so you can watch it; in the app the one-shot clips (wave, focus, done) play once and hand over on their last frame. These concept clips are 480p. The app set should be rendered at 720p.
- Loops run until the state changes. The switch cross-fades over 150 ms. Idle moves so little that the fade from any idle frame to the home frame isn't visible.
- One-shots hand over on their last frame, which is the next clip's first frame, so there is no cut to see.
- The camera and the dunes never move. A cross-fade only ever blends Fen, never the whole picture.
- Stills cover everything that can't play video: reduced motion, Low Power Mode, widgets, the watch and notifications.
- Flutter plays them with
video_player, keeping three decoders warm at most (current, next, idle). H.264 for Android, HEVC for iOS. These 480p clips are about 200 KB each; the 720p set should stay well under 1 MB a clip. - Nothing is lip-synced. Fen doesn't talk; he listens and writes, which keeps every clip reusable.
Clips still to render for v1
The app
Screens
Fen takes a third of the home screen and shrinks to a quarter while recording, so the words get the room. The page under him is his notebook: warm paper, plum-ink text, one apricot button. Each speaker keeps one colour everywhere: dots, labels, owners and the waveform.
2 transcripts mention "van"
"I'll book the van for six in the morning."
"…if the van is late we lose the lift booking."
We sold out by eleven last time. Let's go up to sixty sourdough and forty rye.
And the cinnamon buns were the hit. Priya, can you price a box of six?
Twelve dollars for six feels right. I'll make a little sign for it.
I'll book the van for six in the morning.
Okay, quick one before we open. I'm Maya, for anyone new.
Right. The organisers want our stall list by Thursday.
Who's who?I guessed from what people said. Fix anything I got wrong.
Beyond the phone
Watch and widgets
Apple Watch
Record02:14Sent to your iPhone.Fen's writing it up.
- The watch records; the phone does the work. Audio is saved on the watch and sent with WatchConnectivity file transfer, the same path as Voice Journal's watch app. The phone runs the clean copy and summary.
- Phone not nearby? In v2 the watch uploads straight to our Worker over Wi-Fi or LTE.
- No live text on the wrist. The screen is too small to read a conversation, and streaming from the watch drains its battery.
- Complications: a record button, and "3 to-dos" from the last summary.
Widgets
Fen
- Widgets show stills; WidgetKit and Android widgets can't play video. Fen's pose follows the app: looking at you, happy after new notes, napping at night.
- Small: Fen and a record button. Medium: the last summary, its to-dos, Record and Open. Lock screen: a record button and the next to-do.
- Faster than a widget: a Control Center button and an Action Button shortcut through App Intents start recording without opening the app.
- Android gets the same small and medium widgets.
Under the hood
Two passes, one Worker
The live pass gives you words while people talk. The second pass runs on the saved audio when you stop. AssemblyAI's docs say a streaming speaker label can't be revised once it is given, so the clean copy is where who-said-what gets fixed. Both API keys live only in our Cloudflare Worker.
While you record
- App → WorkerGet a short-lived streaming token. The Worker checks the plan's minutes and asks AssemblyAI for a temporary token, so the real key never ships in the app.
- App → AssemblyAIStream 16 kHz mono audio over the WebSocket with
speaker_labels: true. The same audio is written to a file on the phone. - AssemblyAI → AppTurns arrive labelled A, B, C. The app shows Speaker 1, 2, 3, adds a chip when a new voice appears, and plays Fen's writing loop.
- AppClose the socket on pause. Streaming is billed for every second the socket is open, silence included.
When you stop
- App → WorkerUpload the audio file. If the phone is offline, it waits in a queue and Fen shows a note.
- Worker → AssemblyAIClean copy: Universal-3.5 Pro with speaker labels and speaker identification, which guesses names from the conversation.
- Worker → OpenAISummary: a small GPT-5-class model returns title, gist, decisions and to-dos with owners as structured JSON.
- AppSave, then back up. The clean copy replaces the live text on the phone; the text is backed up to the account.
| Cost of one recorded hour | What it buys | Per hour |
|---|---|---|
| Universal-Streaming (English or multilingual) | Live words while recording | $0.15 |
| Streaming speaker labels | Speaker 1, 2, 3 live, new-speaker chips, Fen's ear flick | $0.12 |
| Universal-3.5 Pro, pre-recorded | The clean copy | $0.21 |
| Speaker labels + speaker identification | Who said what, fixed, plus name guesses | $0.04 |
| OpenAI summaryAbout 15k tokens in, 1k out, small model | Title, gist, decisions, to-dos | ≈ $0.01 |
| Total | ≈ $0.53 |
AssemblyAI list prices on 26 Oct 2026. New accounts get $50 of free credit, roughly 90 recorded hours of this full pipeline. A leaner option drops the live speaker labels ($0.41 an hour), but then speakers only appear after you stop and Fen can't react to a new voice.
- Saved on the phone first. Transcripts, summaries and audio live in a local database, so the library works offline.
- Account rules match the other Bitfuzed apps. Signing out backs up and then clears the phone; deleting the account wipes the cloud and the phone, after confirming who you are.
- Hybrid subscriptions. Buy without an account; a purchase made while signed in follows the account to Android and iPhone.
Your call
Decisions before the build
The look, the neckerchief, the morning sky. Changes now cost cents; after the 720p clip set they cost dollars.
One Pro hour costs about $0.53, so the monthly cap must stay below what the store pays us. Otter Pro is $16.99 a month.
Straw man: Free 30 min a month. Pro $9.99 a month or $79.99 a year, 10 hours a month. Even a yearly subscriber at the cap stays in profit after Apple's 15%.
Voice Journal never uploads audio. Here, audio lets you replay and re-run a transcript on a new phone, but it is the costly part to store.
Suggest: text backup for everyone; audio backup for Pro.
The clean copy covers many languages; live streaming covers fewer. The app UI also has no translations yet, like our other apps.
Suggest: English first, with the multilingual streaming model already wired in.
Lectures run an hour or more. The phone keeps recording with the screen locked through a microphone foreground service (the Voice Journal fix).
Suggest: Free 10 min per recording, Pro 3 hours.
An AssemblyAI account and key, and the OpenAI key, both go into Worker secrets. Nothing is needed from you until the Worker is built.
What happens next
- You approve FenOr tell me what to change: ears, colours, scene, pose.
- Render the app clip setTen clips at 720p, about $18 on Higgsfield.
- App skeletonFlutter app, the fox player, recording, local library, search.
- The WorkerStreaming tokens, upload, clean copy, summary, plan limits. Needs the AssemblyAI key.
- Watch and widgetsSwiftUI watch recorder, WidgetKit and Android widgets, Control Center button.
- Accounts and billingSign-in, backup, hybrid subscriptions, store listings.