On the phone, private, and built to be read all evening.

Ozen (the word means "ear") turns what the people around you say into large, steady captions on your phone, as they speak. It was built for one grandmother with hearing loss, who found every existing transcription app either unreliable with external microphones or too shallow to trust.

• Captions that hold still. Words settle in place instead of jumping around, and a name shows who is talking.
• Private by design. No account and no company server: nothing leaves the phone unless you choose to send it.
• Alerts she won't miss. The doorbell, a smoke alarm or her own name makes the phone buzz, even in a pocket; alarms flash the whole screen.
• Talk back. Type a reply and the phone says it out loud, or shows it in huge letters.
• Faster with a home computer. Optionally, a PC with a graphics card writes the captions, faster and more accurately than the phone.

Website: ozen.live, in English and Hebrew.

Get it: each release has Ozen.apk for Android phones and Ozen.ipa for iPhone and iPad; see Installing on Android and Installing on a phone without a Mac.

Note

The Android app is new and still a little buggy. It is in its first round of fixes: on some phones text can overlap or sit out of place, and those are being found and fixed now. It passes its tests on an emulator, but nobody has used it on a real phone yet. If something looks wrong, please open an issue.

Guides: Getting started · Where your words go · Cloud captions: which service? · Captions from a computer at home · Helping someone who uses Ozen

How it works

Gallery

Live captions, with who said each line. A phrase she asked to hear about lights up.	A light theme for daytime reading.	Type a reply and the phone says it aloud. Common phrases are one tap away.
Or show it full screen, in letters anyone across the table can read.	Settings in plain words, with a live preview of the text size.	Past conversations are saved on the phone, and can be deleted any time.

_{Shown in English, the sample conversation too, so it reads here. Ozen is built around Hebrew speech and can caption any of its twelve interface languages instead, chosen on first launch or in Settings.}

Why this exists

Existing apps had three concrete problems: inconsistent handling of external microphones (no explicit input picker, a USB-C lavalier mic didn't work at all, AirPods worked but not reliably), live transcription treated as a minor feature rather than the point of the app, and no real speaker detection. Ozen exists to fix specifically those three things, for Hebrew conversation, on the phone by default.

What it does

Live, steady captions Words appear as they're said and settle in place once they're stable.	Made for reading all evening Text from 20 to 64 pt, three colour schemes, and numbers that stand out.	Who said what Voices are told apart on the phone, and named once you record them.
Any microphone The phone's own, Bluetooth, wired or USB-C, picked up again when it reconnects.	Sounds she can't hear Her name, the doorbell or a smoke alarm buzz the phone; alarms flash the screen.	On the lock screen The newest two lines, readable without unlocking the phone.
Replies, aloud or huge Type a reply and the phone says it aloud, or shows it in huge letters.	History and stars Saved, searchable conversations, with the important lines starred.	Stays on the phone No account and no company server: nothing leaves the phone unless you choose to send it.
Recovers by itself A dropped recognizer, a failed audio session or a stalled download is retried, and the status line says so.	Home computer, optional A PC with an NVIDIA card can write the captions instead, paired with a QR code.	Twelve languages Buttons and settings follow the phone's language or the one picked in Settings. The captions are of the language people around you speak, Hebrew unless you pick another.

Everything it does

Open a section for the details.

Captions: live text that holds still, the speech models, the display

• Live captions, not batch transcription. Partial text updates continuously as speech happens; text only locks in (stops changing) once it's actually stable — see CaptionStabilizer.
• A Whisper trained on Hebrew, then on the room. The recommended model starts from ivrit.ai's continued training of Whisper large-v3-turbo on about 5,000 hours of Hebrew (Apache-2.0). On 90 Hebrew test clips that model got 6.6% of words wrong on lecture speech and 21% on read sentences, against 9.7% and 31% for OpenAI's Turbo and 27% and 48% for Small, all measured before conversion. Ozen then trained its listening half further on Hebrew played through real rooms, in household noise and beside a talking TV (October 2026). On 700 test sentences heard that way it gets 20.1% of words wrong where ivrit.ai's gets 23.1%; on clean speech 13.3% against 13.1%, and on lectures, read sentences and WhatsApp voice notes the two are within half a point (also measured before conversion). It is converted to WhisperKit's format and compressed to 8 bits (819 MB); compressed the same way, ivrit.ai's model wrote the same words as the full one on five clips checked on a Mac, but for gigahertz run together and one verb form. Nobody publishes these in WhisperKit's format, so it is downloaded from the releases of the public ozen-models repository (scripts/model-release/) and compiled on the phone.
• Two swappable on-device engines, plus the cloud — WhisperKit (Whisper via CoreML) and Apple's own on-device Speech framework — picked in Settings, since which one is actually better for Hebrew on a given device is an open, testable question rather than an assumption.
• Whisper model manager: eleven models from 76 MB to 3 GB with honest Hebrew-quality and speed ratings, download progress, disk usage, and delete. "Turbo Hebrew, noise-trained (Ozen)" is the recommended pick and what a fresh install gets; a phone set up when Small or OpenAI's Turbo was the default is offered the switch on the caption screen, one tap and one download. One on ivrit.ai's Turbo keeps it until switched in the list: the step is real in noise but level up close, too small to interrupt the captions for.
• Display built for reading all evening: text size 20–64 pt (pinch the captions to change it), white-on-black / yellow-on-black / black-on-white, or following the phone's own light/dark setting, bold, speaker names on/off (shown once at the start of each person's turn, like a chat), a small question mark on lines the engine itself was unsure of (so she knows when to ask again; with ivrit.ai's models one or two lines in a hundred, and in tests 50 of 52 marked lines had a word wrong; Ozen's noise-trained model is a little surer of itself, so its lines get the mark a little sooner: of 3,500 test lines, near and far, it marked 282 (ivrit.ai's, 297) and 278 of them had a word wrong), numbers (the time of an appointment, how many pills, a phone number) in a heavier weight and a second colour, whether written in digits or in words ("and six" (u-ve-shesh), "three pills" (shlosha kadurim), "twice" (pa'amayim), "once a day" (pa'am be-yom), "in two weeks" (be'od shvu'ayim)), long stretches of speech broken into short paragraphs at sentence ends, screen stays awake while listening (and locks as usual after a quarter hour with nothing said, while captions and alerts carry on), auto-scroll that stops when you scroll up to re-read (with a "To the last line" (la-shura ha-achrona) pill that turns into a count of the lines said meanwhile, like "3 new lines"). The screen keeps the newest 300 lines (while you read back, the lines above you stay put, up to 600); above them a note opens the saved conversation at the last line that scrolled away. With VoiceOver on, each finished line is read out or sent to a braille display by itself, once (again only if a slow engine then corrects its words), with the speaker's name when the speaker changes and, on a line with the question mark, "May not have been heard correctly" first. Doorbell, alarm and name alerts are read out the moment they happen, and so are captions stopping because of a problem and coming back after one, saving failing and a microphone disconnecting.
• Keeps its cool. Whisper refreshes the in-progress line less often when the phone runs hot or Low Power Mode is on, instead of throttling and falling behind; the careful end-of-sentence pass is never skipped.

Voices and microphones: any microphone, who is speaking, the names to listen for

• Explicit microphone selection, including external Bluetooth/wired/ USB-C inputs, with automatic recovery when a preferred input reconnects mid-conversation.
• Speaker detection, always on, with optional voice enrollment so a person's turns are labeled by name. Several 30-second recordings of one person, made in different places and saved under the same name, work better than one long one: on 16 Hebrew speakers in groups of four, three short separate recordings named the right person 84% of the time against 64% for one 30-second recording, and past about a minute of recording, length alone stopped helping. Unnamed voices are numbered from 1 in each conversation, so a phone listening all week doesn't reach "speaker 140". Voices are told apart by WeSpeaker CAM++, a small (7.3M-parameter) neural speaker-embedding model run entirely on-device via CoreML. Lines with no voice in them (a pot put down, a running tap) are skipped before Whisper sees them, using Silero VAD in the CoreML conversion by FluidAudio (MIT). On a LibriSpeech clustering test it separates a simulated four-person table correctly 96.8% of the time, against 45.8% for the classic MFCC print it replaced. A conversation is harder: on 18 put together from 16 Hebrew speakers' recordings, two to four people each, the app's own speaker code labelled who said a sentence right 79% of the time up close and 64% from across a room, and up close it showed more voices than there were (two people came out as three to six).
• Saved speakers can be renamed, and the new name follows onto lines already on screen, into the names list and into every saved conversation, so fixing a misspelled name also fixes last week's conversations and a search for the new name finds them.
• Names and words list. Family names, the doctor, the medicines. Every engine is primed with the list (Apple's recognizer via contextual strings, Whisper via a decoder prompt, the home computer and each cloud service in its own way; on a long list, Whisper on the phone gets only the start of it, about 20 first names or a dozen full names and medicines, so put the important ones first), edits apply from the next sentence, and Whisper output that is just the list read back is dropped.
• Robust audio: a live level meter per microphone, automatic recovery from phone-call interruptions and route changes (AirPods in/out, USB mic unplugged), Whisper-hallucination filtering on silence, no model runs on pure silence at all. A glitching microphone's broken samples are turned into silence before anything hears them, so one bad moment can't leave the speech detector deaf or a voice print unmatchable, and the diagnostics report counts them.

Alerts: her name, the doorbell, a smoke alarm, the battery

• Keyword alerts. Her name, or any word she picks, buzzes the phone and highlights the line, matching through Hebrew's attached prefixes ("ve-le-Ruti" still matches "Ruti"), without firing a short name on the everyday word it hides in ("Li" stays quiet on "sheli", mine; "Ben" on "lavan", white; "Sarah" on "kshera", kosher; "Leah" on "mele'ah", full), nor on a word still being said ("Tal" stays quiet on "ha-tel..." until it turns out to be "ha-telefon", the phone). The suggested "medicine" and "doctor" also match "the medicines" and a woman doctor. Said over and over at the table, it buzzes at most once every 15 seconds, while every line it's in stays highlighted.
• Sound alerts. Doorbell, knocking, a baby crying, a smoke alarm, a civil-defence siren and about 45 more, recognized on the phone by Apple's sound classifier and shown as a banner, with per-sound muting. Sirens and alarms flash the edge of the whole screen for a few seconds, the doorbell and a crying baby twice, so they're caught from the corner of the eye (slower than any seizure risk, and a steady glow with Reduce Motion on). The phone vibrates differently for each, so they can be told apart in a pocket: long buzzes for an alarm, a double knock for the door, three quick taps for her name. The banner and flash also show over whatever screen is open, such as the keyboard for typing a reply. A lesser sound heard meanwhile (a kettle during a smoke alarm) still buzzes but doesn't take the alarm's banner or flash away. While captions wait to try again (the home computer asleep, the cloud out of reach, a model download being retried or waiting for Wi-Fi or for room on the phone) the microphone stays on for sound alerts alone, so a smoke alarm at night isn't missed because the computer was off.
• Alerts reach her with the screen off. With the phone in a pocket or locked, a sound alert or her name becomes a phone notification (once per 30 seconds per sound or word), so the doorbell isn't missed just because nobody was looking at the app. If captions stop there and nothing will bring them back (a call ends but iOS keeps the microphone, or recovery gave up), the app first tries to take the microphone back by itself, and otherwise sends one notification saying so, removed again once captions are back. (iOS may suspend a locked app for the length of a call; then this runs whenever iOS next lets the app run.) The same notification says when the home computer stops answering, since waking it is what brings captions back.
• Battery warnings at 20% and 10% while captions run, because hours of listening drain the phone and nobody following a conversation watches the battery icon. With the phone in a pocket they arrive as a notification, since a phone that switches off takes the alerts with it. They also come while captions wait to come back with the microphone on for sound alerts.

Lock screen and coming back: the lock screen, what she missed, nothing lost

• Captions on the lock screen. While captions run, the newest two lines show on the lock screen as a Live Activity (with the Dynamic Island, the newest words show beside the camera while she is in another app, and the two lines when it is held open), each saying who is talking when speaker names are on and the voice is known, so the last sentence can be read without unlocking the phone. A call or a failure keeps it there and says why the lines stopped; if iOS closes the app it says the lines aren't updating rather than showing an old sentence as new, and after a quiet minute it says how long ago the last line was said. A line from before five quiet minutes never sits above a new one. Settings can turn it off, since anyone looking at the phone can read it.
• Picks up where she stopped reading. Captions carry on with the phone locked or another app open; coming back, a line across the captions marks where the ones she missed begin, with how many there are, and a button at the top jumps up to it ("What was said meanwhile" (ma she-ne'emar beinta'yim)). Being away less than 15 seconds, or missing a single line, doesn't move the mark. After five minutes or more with nothing said, the time the talking started again is drawn between the lines, so an old sentence isn't read as the one just before.
• Bottom buttons that get out of the way: while captions follow the newest line on their own, the buttons slide away after a few seconds so they never cover it; a touch brings them back.
• Nothing lost when iOS closes the app. A speech model is one of the biggest things in a phone's memory, so iOS may end the app in the background mid-conversation. Coming back, the empty screen offers the conversation if it was still going in the last 20 minutes, one tap away. When iOS warns it is short of memory and captions are off, the app lets go of the loaded model first, so it is less likely to be the app iOS ends.

Talking back: typed replies said aloud or shown huge, Siri, Control Center

• Type to speak. The other half of a conversation: type a reply, or tap one of the ready-made phrases ("Wait, I didn't catch that" (rega, lo hevanti), "Could you say that again?" (efshar lachzor al ze?), "Please speak closer to the phone" (dabru karov yoter la-telefon), which helps the captions most), and the phone says it aloud, in the voice of the language it is written in (Hebrew letters get a Hebrew voice). The last typed sentence stays under the field, to say again when it wasn't caught or keep as a ready-made phrase in one tap. Captions pause while the phone talks, so it doesn't caption itself, and come back on their own. A full-screen pad in huge letters lets someone type to her where captions can't keep up, or turns what she typed upside down for the person facing her to read. Siri opens it straight away with "kitvu li be-Ozen" ("write to me in Ozen") in Hebrew or "Big text in Ozen" in English, and so can the Action button.
• Siri and Shortcuts. "Hey Siri, start captions in Ozen" (hey Siri, hatchel ktuviyot be-Ozen), "Stop captions" (atzor ktuviyot), and, to have the phone say something aloud, "tagid be-Ozen" in Hebrew or "Say with Ozen" in English; Siri then asks what to say. Each also in the feminine (hatchili, itzri, tagidi).
• A start button in Control Center (iOS 18): "Start captions" (hatchalat ktuviyot) opens Ozen and starts listening in one press. It can also replace the flashlight or camera button at the bottom of the lock screen.

History: saved, searchable conversations and starred lines

• Conversation history. Conversations are saved as they happen, searchable (every word typed has to be in the conversation, in any order and on any line, and opening a result jumps to where they were said), shareable as text (a long one opens with its lines that had numbers in them, and a line with the question mark says "May not have been heard correctly"), can be given a name ("Visit to the doctor" (bikur etzel harofe)), list who took part, and open with a summary: length, how much each person said, speaking pace, longest turn, and every line with a time, an amount or a phone number in it, each a tap from where it was said. Phone numbers in captions, including the 1-700 and 1-800 numbers health funds and pharmacies give out, are a tap from a call (star numbers such as *2700 stay text: iOS won't dial them from a link). A number is only linked when it has the right length for its prefix, so a date, an hour or a digit said next to it is never dialed with it. Holding a line copies just that line, to paste a phone number or an address somewhere else. The list is grouped under "Today" (hayom), "Yesterday" (etmol), the weekday for the past week, and the date before that. The list reads a small summary per conversation, so it opens quickly even after months of daily use, and autosaving never stutters the captions. Old conversations can delete themselves after a week, a month, three months or a year (off by default); starred or named ones are always kept, and a change that would delete something asks first. If the phone fills up and saving starts failing, the caption screen says so, instead of conversations quietly going unsaved.
• Turn a recording into text. A voice note (WhatsApp's too), a recorded lecture or the sound of a video: share it to Ozen from another app, or tap "Transcribe a recording" in History. Ozen reads it a little at a time, so an hour-long lecture fits, cuts it at the pauses, and writes each stretch with the most accurate model already on the phone for the recording's language, even one too slow for live captions (or offers the download if there is none). The screen shows how far it has got, and it can be stopped. The text comes out as timed lines, saved to History under the file's name, and can be shared as text or as subtitles (.srt). Live captions and a recording can't use the model at the same time: Ozen asks to stop the captions first, and they don't start again by themselves until the recording is done. The recording never leaves the phone.
• Star what matters. Hold a caption line to mark it as important (what the doctor said about the pills), copy it, or say who is talking. Stars are saved with the conversation, counted in the history list, marked in shared text, and one button steps through them later. One list gathers every starred line from every conversation, with the question mark on any the engine was unsure of, and can be shared as text. A search result opens at the lines it found, highlighted.

Privacy, the home computer and the cloud: what leaves the phone, and only when you choose

• Nothing leaves the phone unless you choose to send it. No account, no company server. The exceptions are all explicit and off by default: your own home computer, cloud captions above, and a switch to let Apple's recognizer use Apple's servers when iOS has no on-device Hebrew model. Saved conversations and settings (names, voice prints) are part of the phone's own iCloud or computer backup when that is on, like any app's; the sound clips, the diagnostics journal and the models are kept out of it.
• Your own computer as the recognizer (optional). A Windows or Linux PC with an NVIDIA graphics card, at home, writes the captions instead of the phone: the phone sends its microphone to the PC and gets the words back, on the home Wi‑Fi or from anywhere with internet (through an encrypted wss:// address). The PC runs the full-size Hebrew Whisper (ivrit.ai's large-v3, 1.55 billion parameters, 16-bit) with beam search for each finished sentence, and the faster Turbo for the live words: 8.3% of words wrong on Hebrew conversation, against 8.9% for the phone's model, and 22% against 25% when the speaker sits across the room. Words appear about 0.2 s after they are said; a finished sentence settles about 0.9 s after the speaker stops (on an RTX 2080 Ti). Setup is one double-click (Ozen-Home-Setup.cmd from the latest release) and pairing is a QR code; if the PC can't be reached, the phone's own model carries on, once it has been downloaded as a backup (Settings → Home computer → "Download a backup to the phone"; without it captions wait for the PC, start again within about 15 seconds of it answering, and the status line offers the backup; sound alerts keep listening meanwhile). See Home computer requirements and server/README.md.
• Optional cloud captions for when the phone's models are too slow or lose track of several people talking: what is said goes to the cloud speech service chosen in Settings, Soniox (which writes the words as they are said, over one open connection), Deepgram, OpenAI, Groq, ElevenLabs, Google Gemini, Speechmatics, AssemblyAI (both also write as the words are said), or a Google model through OpenRouter, with your own key pasted into Settings (stored only in the phone's Keychain, one per service), together with the names and important words lists so it can spell them. Which one? Soniox: the most accurate of the cloud services on ivrit.ai's public Hebrew leaderboard, and the cheapest, about 12 cents for every hour the captions run. To try the cloud for free, Deepgram: second best on that test, and a new account gets $200 of free credit with no card. Prices, languages, privacy and how to get each key: docs/cloud-services.md. Through OpenRouter, on twelve Hebrew test recordings its more accurate Gemini model got 24% of words wrong (the faster, cheaper one 29%) against 39% for OpenAI's Whisper large-v3 turbo and 60% for Whisper small, and puts each change of speaker on its own line. The phone's Hebrew-trained model was not in that test; on 60 other read recordings it made a third fewer mistakes than OpenAI's turbo (21% against 31%), so the cloud is no longer clearly the more accurate choice. Except on the services that keep one connection open (Soniox, Speechmatics, AssemblyAI), a sentence is sent again every few seconds while it is said; through OpenRouter the faster model costs about 15 cents per hour of speech, the more accurate about twice that, and unbroken talk like the news up to three times as much; off Wi-Fi that is a few hundred MB of mobile data an hour, up to about 1 GB. See CloudSpeech and CloudStreamEngine. If the key stops working, the credit runs out or the internet goes, a Whisper model already on the phone takes over and the status line says so (CloudCover); captions don't stop for a billing problem.

Reliability: recovery, downloads, the status line, the seven-day install

• Recovers by itself. A recognizer that drops out mid-conversation, an audio session that fails, or a model download that hit a dead Wi-Fi is retried automatically with growing delays, and the status line says so. Failures only a person can fix (a denied permission) never are. Retries wait out a phone call instead of burning attempts during it. A microphone that silently stops delivering audio (a Bluetooth hearing aid reconnecting, say) is noticed within seconds and restarted, instead of the screen saying "listening" over captions that never come.
• A status control that always says what's happening — asking for the microphone, downloading the model (with a percentage and the time left at its current pace), loading it, listening, paused, or exactly what failed and what to do about it (retry, open Settings, switch engine). First-launch model download is the slowest thing the app ever does and is never silent.
• Downloads that can't get stuck. A cut-off model download is told apart from a whole model and simply continues from where it stopped, instead of failing to load on every launch.
• Downloads that don't eat the data plan. On cellular or in Low Data Mode a model waits for Wi-Fi, says how big it is, and starts by itself when Wi-Fi arrives. "Download now" asks first; a setting allows it always.
• Downloads that fit. Free space is checked before a model download starts, with room left for the first load. A phone that's too full says how much to free (or to pick a smaller model), is never retried on a timer, and starts the download by itself when she comes back to the app with enough room. The model list marks the models that won't fit, and they can't be picked over the one that works. Switching to a model that still has to download while captions run asks first, since captions stop until it arrives.
• Warns before the install runs out. Installed with a free Apple ID, the app stops opening after seven days without a word. It reads its own provisioning profile, says on the caption screen two days ahead when that will happen ("Ozen will stop opening tomorrow at 07:24" (Ozen tafsik lehipatach machar be-sha'a 07:24)), and sends a reminder notification the day before, in daytime.

Setup and help: the first launch, twelve languages, diagnostics

• First-launch walkthrough in large type that explains the engines and the one-time model download before it happens, asks for the microphone with a reason, and asks for her name (with "Grandma" (savta) one tap away), so the name alert works from the first conversation instead of waiting for someone to find it in Settings. A phone set up before that page existed gets the same question as a card on the empty caption screen, until a word is added or it's turned down.
• Twelve interface languages (Hebrew, English, Arabic, Russian, Amharic, French, Spanish, Ukrainian, German, Portuguese, Simplified Chinese, Hindi), following the phone's language or a choice in Settings. Every piece of text is written as tr(hebrew, english), with every other language drawn from a translation table keyed by the English text (scripts/check-translations.py checks both); Hebrew and Arabic read right to left, and each caption line reads the way its own words do.
• The spoken language, chosen in the first-launch walkthrough and in Settings apart from the interface language, from the same twelve. Hebrew keeps Ozen's Hebrew-trained model; the other eleven use the general Whisper Turbo model from its public page on Hugging Face (on Android, whisper.cpp's version, downloaded the first time captions start), and a home computer loads the general model while such a phone is connected. The suggested alert names are offered as said in the spoken language, and a lone English "you" is kept when the model is sure of it. Amharic captions are weak, as the general model handles it poorly, and tappable phone numbers are Israeli formats.
• Diagnostics screen with every pipeline counter (audio chunks, tokens, caption lag, restarts, speaker clusters), free space, memory use, a timeline of the last failures, retries, microphone stalls, phone calls and low-memory warnings with clock times, and one-tap copy of all of it for asking for help.

Training a model that hears across the room

Ozen's own version of the speech model, trained to hear better from across a room and through household noise, shipped on 9 October 2026 and is the recommended model: on 700 test sentences heard across a room, in household noise or beside a TV it gets 20.1% of words wrong, against 23.1% for ivrit.ai's Turbo it was trained from, and clean speech is level (13.3% against 13.1%). The models will keep improving over time. Each new version sits the same fixed listening exam and replaces the current one only when it does better. Training ran on a rented RTX PRO 6000 at $1.99 an hour, paid from free GPU credit GarageFarm gave the project. The scripts are in training/.

Home computer requirements

Only for the optional home computer; the phone works on its own.

	Minimum	Recommended
Graphics card	NVIDIA GTX 16-series or RTX 20-series or newer, 6 GB (e.g. RTX 2060)	NVIDIA with 8 GB or more (e.g. RTX 3060 12 GB, 4060, 2080 Ti)
What runs	The fast Turbo model for everything	Turbo for live words plus the full large-v3 for finished sentences (more accurate, most of all from across the room)
Disk	7 GB free	12 GB free
System	Windows 10/11 or Linux, with the normal NVIDIA driver	same
Connection	The phone reaches it on the same Wi‑Fi, or from anywhere through the free Tailscale Funnel the setup offers	same

Older NVIDIA cards (GTX 10-series and earlier) have enough memory but lack the fast 16-bit arithmetic the server uses, and AMD, Intel and Apple graphics aren't supported. Memory in use is about 6.5 GB with both models: 1.6 GB for Turbo and 3.1 GB for large-v3 at 16 bits, the rest the work space of five-way beam search over 30-second windows and NVIDIA's libraries. The Windows setup checks the card and the free space before it downloads anything, and adds the full model on a card with 8 GB or more. On Linux server/setup.sh checks neither, and the full model is added by hand (see server/README.md).

Measured on rented cards (2026-09-27, the server's exact work on the same 100 Hebrew sentences: the full model with beam 5 for each finished sentence, Turbo for live words; every card wrote the same text):

The same numbers as a table

Card	Finished sentence	Live words
RTX 4090	0.33 s	0.12 s
RTX 3090	0.52 s	0.22 s
RTX 4060 Ti	0.69 s	0.27 s
RTX 3060 12 GB	0.77 s	0.34 s
RTX 2080 Ti (Linux)	0.79 s	0.27 s
RTX 3070	0.93 s	0.42 s
RTX 2080 Ti (Windows)	1.02 s	0.23 s

The two 2080 Ti rows are different computers, so the gap between them is a hint that Windows adds overhead to beam search, not a clean measurement.

A faster card shortens only the final pass: words already appear about 0.2 s after they're said, and a sentence counts as finished after a 0.7 s pause whatever the card. Beam search wider than 5 measured no more accurate (beam 20: same 8.3% word error, 1.5 s per sentence instead of 0.9 s); Settings → Home computer → "Speed or accuracy" trades a little accuracy for about 0.15 s at beam 1.

Repo layout

Package.swift          OzenKit: platform-independent core logic
Sources/OzenKit/        (builds + tests on Linux — no Mac needed for this half)
Sources/OzenPlatform/  WhisperKit/Speech/AVFoundation/Accelerate integration
                        (Apple-only; built and tested via CI's macOS runner)
Tests/                  Unit tests for both of the above
App/Ozen/               The SwiftUI app itself (generated via XcodeGen)
App/OzenWidget/         Widget extension: lock screen captions, Control Center button
App/Shared/             Code compiled into both the app and the widget extension
project.yml             XcodeGen config — run `xcodegen generate` to get Ozen.xcodeproj
docs/superpowers/specs/ Design doc with the full rationale and open questions
docs/design/            Design docs for newer work: the Android app, the cloud services
android/core/           The Android app's portable core in Kotlin (tests run on the JVM)
android/app/            The Android app: microphone, whisper.cpp, background service, screens
training/               Training the speech model to hear across a room, and its exam

Building

This was developed without access to a Mac. The portable core is built and tested directly:

swift build && swift test

The full app (anything touching WhisperKit/Speech/AVFoundation/SwiftUI) only builds on iOS — AVAudioSession in particular doesn't exist on macOS at all, so this half is built and tested by CI (.github/workflows/ci.yml) against the iOS Simulator on GitHub's free macOS runners, not a plain macOS build. To build it yourself on a Mac:

brew install xcodegen
xcodegen generate
xcodebuild -project Ozen.xcodeproj -scheme Ozen -destination "platform=iOS Simulator,name=iPhone 16" test

The Android app needs Java 21 and the Android SDK, plus the NDK for whisper.cpp:

cd android
./gradlew :core:test            # the portable core, on the JVM, no emulator
./gradlew :app:assembleDebug    # the debug APK

Installing on Android

Each release has Ozen.apk, built by .github/workflows/release.yml and signed with Ozen's own key, for Android 8 or later on a phone with a 64-bit processor:

1. On the phone, open the latest release and tap Ozen.apk to download it.
2. Open the downloaded file. Android asks whether to allow installing apps from the browser or the Files app; allow it, then tap Install.
3. Open Ozen. The walkthrough asks which language the people around you speak. The first time captions start, Ozen downloads the speech model for that language (574 MB, over Wi-Fi unless "Download models on cellular data too" is on in Settings) and shows how far it has got; after that it works without the internet.

A newer Ozen.apk installs over the old one and keeps the conversations, names and settings. The Hebrew model is Ozen's own, the same one the iPhone uses, converted for whisper.cpp (a release of ozen-models); the other languages use whisper.cpp's general Turbo model from Hugging Face.

Installing on a phone without a Mac

.github/workflows/release.yml builds an unsigned Ozen.ipa on GitHub's macOS runners and attaches it to a release. Signing and installing it with a free Apple ID from a Linux machine is documented step by step, including the three things that were broken along the way, in docs/sideloading-from-linux.md.

On iOS or iPadOS 26 and later, install with Impactor, not AltServer. A build signed by AltServer-Linux installs without an error but never opens: the system rejects its signature before the app starts (AMFI: code signature validation failed in the device log), so it looks like the app crashes on launch and leaves no crash report. Confirmed on iPadOS 27 on 2026-09-27.

When she calls with a problem, docs/troubleshooting.md says what every status message means and what to do about it, and how to get the diagnostics report sent over.

Status

The whole pipeline — permission, session, input listing, engine preparation with progress, capture, tokens → segments, speaker clustering, engine hot-swap, pause/resume, automatic recovery, every failure path — lives in OzenKit as CaptionPipeline and is unit tested on Linux against fakes, along with the alert matching, history, statistics, vocabulary, model-download, recovery, battery, notification and layout logic (over 1,200 tests). The platform layer (WhisperKit/Speech engines, real audio capture, the speaker embedder) and the app's view model are built and tested on CI's iOS Simulator, with the view model driven end to end by the same fakes (about 180 more; the Simulator run, which repeats the portable ones, has about 1,400). The app has been installed and run on a real iPhone 15 Pro Max (iOS 18) and an iPad (iPadOS 27). Actual Hebrew transcription quality, external-mic behaviour and speaker separation in a real room are being verified by hand — see the design doc's checklist. Ozen's own version of the speech model, trained to hear across a noisy room, has been the recommended model since 9 October 2026; see Training.

The Android version, a full Kotlin rewrite started on 10 October 2026, ships in every release since build 119 as Ozen.apk (see Installing on Android). The portable core and the caption pipeline are ported (about 1,700 tests on the JVM), and whisper.cpp runs on the phone with the same Hebrew model as the iPhone, or with its general Turbo model for the other spoken languages, either downloaded the first time captions start. The walkthrough, captions, History, Settings, Diagnostics and typing a reply for the phone to say are checked by about 250 tests on an Android emulator, with sound alerts recognised by Google's YAMNet (Apache 2.0) and speaker names from the same WeSpeaker CAM++ model as the iPhone, both run on the phone with LiteRT. It has not been used on a real phone yet, so expect rough edges; a Bluetooth headset's microphone probably isn't picked up yet (nothing starts the headset's call audio). The plan and how far it has got are in docs/design/2026-10-10-android.md.

Support Ozen

Ozen is free and will stay free, with no ads and nothing locked. Donating is entirely optional: please never feel obligated, and only donate if you are in a financial position where you can comfortably afford it. Using Ozen and telling others about it already helps a lot.

How to donate:

1. Open your crypto wallet app and choose Send.
2. Pick the same coin as listed here. A coin sent to another coin's address is lost.
3. Scan the QR code with the wallet, or copy the address and paste it.
4. Before sending, check that the first and last few characters of the address match.

Bitcoin (BTC)

Ethereum (ETH) plus USDC and USDT on the Ethereum network only

Monero (XMR)

Bitcoin (BTC)

bc1qk5aym0mch042200s2wrc366r3hsxxmgc9nu7tm

Ethereum (ETH, plus USDC and USDT on the Ethereum network only)

0x0Ea2210fcB0BbF2C3202d9663dB762F1f51b1BBC

Monero (XMR)

46otohcpNKQfFi9F21ZHTcSiNVrLMw4yMS1SFM5hbDfu5LZCzLGkEZ2Vx4YD5kwK3nKUG6GjMf37z7i6sFQR2NEC1W9ubhb

The same addresses and codes are in the Android app under Settings, About, Support Ozen. The iPhone app has no donation page, as the App Store doesn't allow one outside its own payments.

In the press

• Breaking the Sound Barrier: How This Developer Made Accessible Conversations, Xesktop blog, 16 October 2026: why Ozen exists, and how its Hebrew speech model was trained on Xesktop's GPU servers.

Supporters

These companies gave or offered Ozen free plans or credits. Thank you.

• AI code reviews provided by CodeAnt
• AI code reviews provided by CodeRabbit
• Security scanning by Snyk
• Audio processing support from LALAL.AI
• GPU training credits provided by GarageFarm.NET
• Voice AI credits from Retell AI
• Audio processing support from Auphonic
• Database tooling support from Prisma (credit offered)
• GraphQL engine support from Hasura (credit offered)
• Package hosting support from Cloudsmith (open-source plan under review)
• LLM observability support from Langfuse (credit offered)

License

GNU Affero General Public License v3.0 — see LICENSE.

Ozen is free software, and every copy has to stay free: anyone who shares the app, a changed version of it, or runs a changed home-computer server for other people must pass on the full source code under the same license. Code taken from an earlier version before 2026-09-28 was under MIT; everything from then on is AGPL-3.0.

Made by Arbel.