LEGENDA Captions · Translation

Legenda: subtitles for the conversation you are in

A missed word should not cost you the next sentence. Put your iPhone on the table. Read what is said, with a translation below when you want one.

Johannes Klumpe · Technical note · November 2026

Abstract. In a language you are still learning, a live conversation can move faster than you can follow. Asking for a repeat interrupts it. Guessing risks an error. Staying silent means dropping out. Legenda puts live captions on the table, with optional translation. You can read back, then return to the live line. The transcript becomes your notes. After setup, speech recognition and translation run offline on the iPhone. You control what leaves it. Six languages can share one conversation. This note sets out the evidence and the limits.

Runs on
iPhone · iOS 26 or later · portrait and landscape
Device tested
iPhone 18 Pro only
Input
The iPhone microphone
Output
Live captions, optional translation, plain-text transcript
Network
Needed for setup and new translation pairs; offline afterwards
Status
Private beta on TestFlight. Not on the App Store yet.

1 PremiseEvery way of recovering a missed word has a cost

A non-native speaker has to understand the sentence and keep a place in the conversation. Losing one word can make the next sentence harder to follow. The usual repairs ask something of the speaker or the listener.

ResponseWhat it gives youWhat it costs
Ask for a repeatAnother chance to hear the wordsThe conversation stops for the repair
GuessA way to keep goingA wrong word can change the meaning
Stay silentTime to listenYou lose your part in the exchange
Read LegendaCaptions beside the conversationA glance at the phone; captions can be wrong

Table 1. Common responses to a missed word. These are tradeoffs, not a claim that every conversation follows the same pattern.

Legenda adds a written path back into the conversation. The name is the Portuguese word for subtitle. The app on the phone is called Legenda.

Room: people talking

↓ sound enters

iPhone · processing after setup

Microphone
↓ audio in memory
Voice detection · Silero VAD
↓ speech
Speech model · Parakeet on the Neural Engine
↓ text
Captions
↓ optional
Translation · Apple Translation
↓
Screen · captions, plus translation when enabled

↓ Share · only when you tap it

Your chosen app receives the transcript

Fig. 1. The conversation path after setup. Every processing box is inside the iPhone. The outgoing arrow is your choice to share. Copying the transcript is also explicit. Setup downloads are described in section 5.

2 Claim 1You can follow the words without asking the speaker to stop

The phone lies on the table and listens through its microphone. Captions show what is said. Turn on Translate to split the screen: captions above, translation below in yellow. One menu selects the target language.

Translation works on finalized caption lines. A line ends at a pause or at a length limit, so it can end in the middle of a sentence. The translation therefore trails the caption. It can also wait for a language download or for earlier lines. A line already in the target language appears unchanged. The screen stays awake while the app listens.

Fig. 2. Captions alone, optional translation, and the target language menu. All screenshots on this page show demo text. Tap an image to enlarge it.

In the author's field logs from one iPhone 18 Pro, a recognition pass took about 0.2 to 0.3 seconds. Captions updated about every 1.5 seconds. A pass time measures recognition work, not the delay from speaking a word to seeing it.

3 Claim 2A missed line stays within reach

Scroll up to read again what was said a minute ago. New lines do not pull you back while you read older captions. Tap Live, or scroll to the bottom, to return to the live line. Each pane scrolls on its own.

Your actionWhat the view does
Read older captionsKeeps your place as new captions arrive
Read older translationScrolls independently of the captions
Return to LiveFollows the newest line again

Table 2. The scroll rule is the same in portrait and landscape. Display history is limited to the most recent 200 caption lines; the export has a separate limit.

Fig. 3. The two panes in landscape. The toolbar moves to the top. This demo shows the live view; Table 2 describes reading back.

4 Claim 3The transcript is the notes

You do not have to write the words down again. Share transcript opens the iOS share sheet for Notes, Files, Mail or another app that accepts text. Copy transcript puts the same text on the clipboard.

The transcript contains the captions, not the translation. It is plain text, one line per caption line. It retains up to the most recent 20,000 finalized lines, plus the current unfinished line when you export. If older lines have been dropped, the export says so.

Fig. 4. The transcript enters the standard iOS share sheet. This demo device shows Copy and Save to Files; destinations depend on the apps available on your phone.

5 Claim 4The conversation runs offline after setup

Speech recognition uses NVIDIA Parakeet, the FluidInference/parakeet-ultra-coreml build, through the open-source FluidAudio library. It runs on the iPhone's Apple Neural Engine. Silero VAD detects voice activity. Apple's Translation framework translates on the device.

First start needs a network connection. Setup downloads about 630 MB for the speech model from Hugging Face, plus about 2 MB for voice detection. It then prepares the speech model for the Neural Engine. iOS downloads a translation language pair the first time that pair is needed. With these assets installed, captions and translation work in airplane mode.

Fig. 5. First-run setup. The 42% progress and 265 of 630 MB are demo values, not a timed download test.
StageObserved time
First preparation1 minute 50 seconds and 5 minutes 10 seconds
Later startsAbout 4 seconds
Preparation after an update1 minute 39 seconds in one observation

Table 3. Observations from the author's field logs on one iPhone 18 Pro. Preparation is roughly 2 to 5 minutes, separate from the download. An app update can require preparation again. These are observations, not timing guarantees.

6 Claim 5You decide when the conversation leaves the phone

There is no account, no sign-in and no server of ours. The app has no analytics or crash-reporting SDK. Its own network code fetches model files and setup metadata. It does not upload the conversation.

Audio is processed in memory and is not saved. The transcript also lives in memory. It is lost when the app process closes unless you have shared or copied it elsewhere. Switching away from the app is not the same as closing its process.

One exception is explicit. A hidden diagnostics panel has the action Export log (includes captions). It writes a temporary text file with diagnostic events and caption text, and opens the share sheet for it. This happens only when you tap it. The app never saves a session by itself.

Pause and continue use one button. Sharing the transcript, copying it, and exporting a diagnostic log are your actions. Tell the people at the table that your phone is captioning the conversation.

Fig. 6. Pausing leaves the captions available to read. The play control continues listening.

7 Claim 6Six languages can share one conversation

Legenda detects the spoken language automatically for each finalized caption line. People can mix Portuguese, English, German, Spanish, French and Italian in one conversation. The same six languages are available as the target language.

LanguageCaptionsTarget languageOur speech test
PortugueseYesYes100 utterances per condition
EnglishYesYes100 utterances per condition
GermanYesYes100 utterances per condition
SpanishYesYesNot measured by us
FrenchYesYesNot measured by us
ItalianYesYesNot measured by us

Table 4. Supported languages and test coverage. The speech test measures recognition, not translation quality.

The speech model was tested on FLEURS read speech: 100 utterances each in Brazilian Portuguese, German and US English. The noisy condition adds babble at 10 dB signal-to-noise ratio and reverberation to the same utterances. Word error rate counts substitutions, deletions and insertions relative to the reference words. Lower is better.

0%4%8%12% Portuguese 3.55% 7.64% German 3.74% 11.35% English 5.11% 6.71% Clean Noisy · 10 dB SNR + reverb
Fig. 7. Parakeet Ultra ANE word error rates. Source: R2 benchmark, 22 November 2026, U-ANE column. Measured on a Mac mini M4 with the Neural Engine and whole utterances, using FluidAudio 0.18.4. This is the same model, but it is not a phone measurement of a live conversation. Spanish, French and Italian were not measured by us.

A separate, small test used synthetic speech with mixed languages. In 29 of 37 checks the expected English word was present in the recognized text. The test checks that the word is present, not where it stands. It is indicative only. Names and loanwords can be wrong.

8 LimitsThe captions remain a reading aid

9 StatusLegenda is in private beta

Legenda is in private beta on TestFlight. It is not on the App Store yet.