Best Whisper Apps for iPhone and iOS in 2026: Which Model Each One Runs

Quick answer: The best Whisper apps for iOS in 2026 are LoroNote (Whisper Large V3 Turbo, on device), Whisper Notes (Whisper plus Parakeet and SenseVoice, on device), Aiko (Whisper medium or small on iPhone, on device), and Whisper Memos (Whisper in the cloud, delivered to your inbox). The label “Whisper app” names the model family only — it tells you nothing about which size of Whisper runs, or whether your audio stays on the iPhone. Those two facts differ across all four apps, and they matter more than the label.

• Best Whisper model on iPhone: LoroNote — Large V3 Turbo, the largest Whisper variant any of these runs locally on iOS
• Best for switching models: Whisper Notes — Whisper, Parakeet, and SenseVoice in one app, $6.99 once
• Best one-time purchase across Apple platforms: Aiko — $24.00 once, also on Mac and Apple Vision
• Best if you want transcripts emailed to you: Whisper Memos — cloud processing, automatic AI summaries

This comparison was compiled by the LoroNote team on August 13, 2026. LoroNote is our own app. Every claim about another app is taken from its own US App Store listing, read on that date, and anything a listing does not state is marked “not stated” rather than guessed.

What is a Whisper app?

Whisper is an open-source speech recognition model family released by OpenAI. A “Whisper app” is any app that uses one of those models to turn audio into text.

That single label hides two decisions that change the product completely:

1. Which Whisper variant. Whisper ships in sizes — tiny, base, small, medium, and large, plus the newer large-v3 and large-v3-turbo. Accuracy climbs with size; so do memory use, processing time, and battery drain. A phone running small and a phone running large-v3-turbo are not doing the same job.
2. Where it runs. Whisper is open source, so an app can run it on your iPhone — or call a server that runs it. Both are legitimately “powered by Whisper.” Only one keeps your audio on your device.

Neither fact appears in the App Store category, and most listings do not volunteer them. Assembling them is what this comparison is for.

”Whisper” does not tell you where your audio goes

This is the most common misreading, and it is worth stating plainly: an app being built on Whisper does not make it private.

Whisper Memos and VoicePen both say they are powered by OpenAI Whisper. Both send your recording to a server to be transcribed. That is a reasonable engineering choice — servers are faster, handle any device, and make features like emailed summaries possible — but it is the opposite of what many people assume when they see the Whisper name attached to an open-source, self-hostable model.

If keeping audio on the device is your actual requirement, the label to look for is not “Whisper.” It is an explicit statement that transcription runs locally, ideally with a claim you can test yourself. Our offline speech-to-text guide covers how that data flow works, and our offline app shortlist ranks apps by that criterion specifically. This article answers a different question: not “does it work without internet,” but “which Whisper is actually running, and how good is it.”

Which Whisper model does each app run?

Everything in this table comes from each app’s US App Store listing, read on August 13, 2026.

Two things stand out. Aiko is the only app that publishes exactly which variant it loads, and it is candid that the iPhone gets a smaller model than the Mac — it “uses the Whisper medium or small model on iOS depending on available memory and the large v3 model on macOS.” And LoroNote is the only one running a Large-series variant on the phone itself.

LoroNote’s worldwide rating across all storefronts is 4.7 from 279 ratings; the table uses US figures so every row shares one basis.

1. LoroNote — the largest Whisper variant running on the phone

LoroNote runs Whisper Large V3 Turbo directly on iPhone and iPad. Turbo is OpenAI’s speed-optimized derivative of large-v3, and it is the reason a Large-series model is practical on a phone at all. The next section breaks down exactly what that buys you and what it costs; we also covered the decision in why LoroNote uses Whisper Large V3 Turbo.

Audio and text are processed locally, so transcription works in Airplane Mode and a confidential recording never reaches a transcription server. Speaker identification also runs on the device — you can rename speakers and the labels survive when you share the transcript. See how speaker diarization works.

Around the model sits the workflow: synchronized highlighting and tap-to-seek playback, folders, calendar, search, batch find-and-replace, and a custom dictionary for names and technical terms. Capture covers in-app recording, Apple Watch recording, and importing MP3, WAV, M4A, MP4, and MOV. Exports are TXT and SRT. On supported devices Apple Intelligence can summarize a transcript or pull action items without sending it anywhere.

Best for: getting the most accurate on-device Whisper result available on an iPhone, in any of 100+ languages.

Keep in mind: a Large-series model needs recent hardware — iPhone 13 or later, iPhone SE (3rd generation), or a supported recent iPad, on iOS 18.6 or later. There is no Mac, Windows, web, or Android version. Bigger model also means longer processing than a small model would take on the same file.

How Whisper Large V3 Turbo actually performs

Turbo is the variant that decides whether a Large-series Whisper is usable on a phone, so it is worth going through in detail. Every figure below is from OpenAI’s own model table and the large-v3-turbo release notes.

What OpenAI changed to make it fast

Turbo is not a smaller Whisper trained from scratch. It is large-v3 with the decoder cut down: 4 decoder layers instead of the 32 in the large series — the same decoder depth as tiny. The encoder, which is the part that builds an acoustic representation of the audio, is left intact.

That asymmetry is the whole trick. Whisper’s encoder runs once over each 30-second window, but the decoder runs autoregressively, once per output token. On a long recording the decoder dominates total time, so cutting its depth by a factor of eight cuts most of the cost while the encoder keeps the multilingual acoustic quality that makes large-v3 good in the first place.

OpenAI then fine-tuned the result “for two more epochs over the same amount of multilingual transcription data used for training large-v3” — deliberately excluding translation data.

Published speed and memory figures

Two numbers deserve emphasis. Turbo is roughly 8× faster than large while carrying just over half its parameters (809 M against 1,550 M). And it sits at a parameter count comparable to medium (769 M) — the variant Aiko loads on iPhone — while retaining the large-v3 encoder. That is the crux of the comparison in this article: Turbo is roughly medium-sized in weight but large-sized in acoustic modelling.

The release notes go further on speed: combined with the accompanying optimization patches, “the ASR speed of turbo is faster than what tiny used to be” — a model 20× smaller.

Accuracy: similar to large-v2, with two known exceptions

OpenAI’s own evaluation is that across languages Turbo “performs similarly to large-v2.” It is not marketed as matching large-v3, and it should not be described that way.

The documented exceptions matter if they are your languages: Turbo “shows larger degradation on some languages like Thai and Cantonese.” Results are also stronger on clean read-speech datasets such as FLEURS than on noisier crowd-sourced audio like Common Voice — a reminder that published WER figures flatter every model relative to a real meeting recording.

What Turbo does not do

• Translation. Because translation data was excluded from the fine-tuning, OpenAI explicitly does not expect Turbo to perform well at speech translation. It transcribes the language that was spoken. If you need speech-to-English translation, large-v3 is the right model and Turbo is the wrong one.
• English-only variant. Unlike tiny through medium, there is no .en build of Turbo. It is multilingual only.
• Beat large-v3 on accuracy. It trades a step of quality for roughly 8× the speed. That is the deal.

What these figures mean on an iPhone

One caveat that most write-ups skip: OpenAI’s ~8× and ~6 GB figures are reference figures for GPU inference, not iPhone measurements. A phone has no discrete VRAM pool, and shipping apps convert and quantize these models to run under Core ML and the Neural Engine, which changes both memory footprint and throughput. Treat the table as the relationship between variants, not as a spec sheet for any particular app.

What does carry over to a phone is the shape of the tradeoff:

• Long recordings benefit most. Decoder pruning saves time per generated token, so a 60-minute lecture gains far more than a 30-second memo.
• Memory is the real constraint on mobile, not FLOPs. This is exactly why Aiko’s listing says it picks medium or small on iOS “depending on available memory,” and why device requirements are stricter for apps running a Large-series variant. LoroNote requires iPhone 13 or later, iPhone SE (3rd generation), or a supported recent iPad.
• Short clips are the worst case, not long ones. Per-file overhead — model warm-up and windowing — is fixed, so it dominates a ten-second memo and disappears into a one-hour lecture. Throughput climbs with file length until thermal limits push back.
• Accuracy differences show up where audio is hard, not on clean dictation — accents, crosstalk, background noise, and technical vocabulary. On a quiet close-mic recording in English, a smaller model often produces text you could not tell apart.

How long Turbo actually takes on an iPhone

The useful metric is speed factor: seconds of audio transcribed per second of processing time. A speed factor of 60 means one minute of audio is transcribed in one second. It is the number that actually answers “how long will I wait,” which raw parameter counts never do.

What LoroNote measures on an iPhone 15 Pro

LoroNote shows its running Speed Factor on screen while it transcribes, so these figures are read directly off the app rather than inferred. Transcribing a 10 minute 31 second recording with Whisper Large V3 Turbo on an iPhone 15 Pro:

That settles at roughly 19× — about 33 seconds of processing for a 10.5-minute recording. These are LoroNote’s own in-app measurements on our device, not an independent benchmark; weigh them as you would any first-party figure.

At that rate:

The practical summary: Turbo gives an iPhone the large-v3 encoder at roughly medium weight and roughly 8× the decode speed of full large, at the cost of a small accuracy step versus large-v3, no translation, and known weakness in Thai and Cantonese. On an iPhone 15 Pro that means an hour-long recording becomes text in around three minutes.

2. Whisper Notes — three models, switchable, one payment

Whisper Notes is the only app here that lets you switch engines: Whisper, NVIDIA’s Parakeet, and SenseVoice, all running on the device. Its listing explains when each wins — SenseVoice is “much faster than Whisper and strongest for Chinese, Japanese, and Korean,” while Parakeet is the speed option.

It is also the only listing that publishes real timing figures: transcribing five minutes of audio on an iPhone 15 takes “~18 seconds with Parakeet, ~1 minute with Whisper.” Those numbers are the developer’s own, but they are specific and testable, which is more than most apps offer. If you want the background on that model rivalry, we compared them in Parakeet V3 vs Whisper Large V3.

Everything runs locally — the listing invites you to “turn on airplane mode before you record and leave it on.” You get speaker labels with renaming, no recording length limit, 100+ languages, and export to SRT, VTT, or TXT with timestamps and speaker labels. It runs on Mac as well as iPhone and iPad, and costs $6.99 once with no subscription or in-app purchases.

Best for: multilingual work where you want to pick the engine per recording, and anyone who prefers a single payment.

Keep in mind: the listing does not state which Whisper size it loads, and it says the models are “optimized for English.”

3. Aiko — the most transparent, and the most limited

Aiko is a $24.00 one-time purchase built on Whisper via whisper.cpp, the C/C++ port by Georgi Gerganov, which its listing credits directly. Nothing leaves the device, it supports audio in 100 languages, and it exports subtitles.

It is the most honest listing in this comparison in both directions. It publishes its exact model selection — medium or small on iOS depending on available memory, large-v3 on macOS — and it states plainly that it “favors accuracy over speed,” does not do live transcription while recording, does not currently have speaker detection, and does not plan in-app text editing. The intended workflow is to export and edit elsewhere.

It is also the only app here that runs on Mac and Apple Vision alongside iPhone and iPad. We looked at it more closely in LoroNote vs Aiko.

Best for: people who want one purchase across Apple platforms and are transcribing finished recordings rather than running meetings.

Keep in mind: it now requires iOS 26 or later, the iPhone gets a smaller Whisper than the Mac does, and there is no speaker detection, live transcription, or in-app editing.

4. Whisper Memos — Whisper, but in the cloud

Whisper Memos inverts the usual workflow: you record, and the transcript plus an automatic AI summary arrive in your email. It offers a choice between OpenAI Whisper and ElevenLabs Scribe, which its listing rates at “up to 96.7% accuracy” and recommends for accents and multilingual speech. It covers over 50 languages and imports MP3, M4A, WAV, AAC, FLAC, and OGG.

Capture is genuinely well built — Lock Screen widget, Apple Watch app with standalone recording, Siri, Shortcuts, and the Action Button. Transcripts route onward to Notion, Reflect, Todoist, Things 3, Day One, Evernote, Trello, Drafts, and thousands more through Zapier.

The listing describes “offline recording — syncs when you’re back online,” which is the tell: recording works offline, transcription does not. Your audio goes to a server. Include it on your shortlist if that is fine for the material you record, and rule it out if it is not.

Best for: capture-and-forget voice journaling and note capture where the transcript should land in your inbox and downstream tools automatically.

Keep in mind: cloud processing, the Whisper variant is not stated, and there is no Mac app.

Also Whisper-powered, but a different job

Four more apps come up in a search for Whisper on iOS and are worth naming so you can rule them in or out quickly:

• WhisperBoard (free, open source) — an iOS app built on whisper.cpp where you browse and download the Whisper model size you want inside the app, and everything runs locally. The trade-off is open-source-project polish rather than a finished product, but it is the cheapest way to experiment with different Whisper variants on your own phone.
• VoicePen (free, 4.7 from 505 US ratings) — transcription and AI notes “powered by OpenAI Whisper,” in the cloud, with YouTube, podcast, and Zoom link ingestion plus Notion, Zapier, and Make integrations.
• V2T: Whisper AI Transcription (free, 4.4 from 222 US ratings) — unusual in offering both: on-device live transcription with no internet required, and cloud transcription for uploaded files on a subscription.
• Superwhisper (free, 4.4 from 810 US ratings) — a dictation app rather than a recording transcriber; you bring your own API keys.

Superwhisper in particular solves the other speech-to-text problem — speaking into a text field instead of typing — which we separated out in the broader iOS speech-to-text comparison.

How to choose

• You want the most accurate result the phone can produce: LoroNote, for Large V3 Turbo on device.
• You record in Chinese, Japanese, or Korean: Whisper Notes, so you can switch to SenseVoice for those languages.
• You need a Mac app and one payment: Aiko, accepting a smaller model on the phone and no speaker labels.
• You want transcripts to appear in your inbox without opening an app: Whisper Memos, accepting cloud processing.
• You are dictating rather than transcribing: none of these — look at a keyboard-replacement app instead.

Then test. Record three to five minutes in your normal environment with your microphone, your accent, and a few names and technical terms. Check missing words, punctuation, speaker changes, and elapsed time. Model size explains a lot of the variance between apps; your own audio explains the rest.

Frequently asked questions

Which Whisper model runs on an iPhone?

It depends on the app. LoroNote runs Whisper Large V3 Turbo on device. Aiko’s listing states it uses the Whisper medium or small model on iOS depending on available memory, and the large-v3 model on macOS. Whisper Notes and Whisper Memos do not state a variant in their listings.

Does a Whisper app work offline?

Only if it runs the model on your device. LoroNote, Whisper Notes, and Aiko all state that transcription happens locally, so they work in Airplane Mode. Whisper Memos and VoicePen send audio to a server, so they need a connection to produce a transcript — Whisper Memos records offline and syncs later.

Is Whisper free to use on iOS?

The model is open source, but the apps price differently. LoroNote and Whisper Memos are free downloads with in-app purchases. Whisper Notes is $6.99 once. Aiko is $24.00 once. Apps that run Whisper locally have no per-minute server cost, which is why several of them charge a single price instead of a subscription.

What is the best free Whisper app for iPhone?

Depends on what “free” needs to include. WhisperBoard is genuinely free and open source, runs Whisper locally, and lets you pick the model size — with the rough edges of a community project. V2T offers a free on-device live transcription path. LoroNote is a free download with in-app purchases, so you can test Whisper Large V3 Turbo on your own audio before paying anything. Free cloud Whisper apps exist too, but their server costs are being paid somehow — check what the free tier caps and where your audio goes.

How much faster is Whisper Large V3 Turbo?

OpenAI’s reference table lists Turbo at roughly 8× the speed of large (which is the 1× baseline), with 809 M parameters against 1,550 M and ~6 GB of VRAM against ~10 GB. The release notes add that with the accompanying optimization patches, Turbo’s ASR speed is faster than tiny’s used to be.

Those are GPU reference figures. On actual Apple hardware, LoroNote’s in-app readout shows a speed factor around 19× on an iPhone 15 Pro for a ten-minute recording — roughly 33 seconds of processing for 10.5 minutes of audio.

How long does it take to transcribe an hour of audio on an iPhone?

On an iPhone 15 Pro running Whisper Large V3 Turbo, LoroNote reports roughly 19× real time, which puts a 60-minute recording at about three minutes. That is a first-party figure read from the app’s own progress display rather than an independent benchmark. Other apps publish lower numbers — Whisper Notes’ listing works out to about 5× on its Whisper path.

Is Large V3 Turbo as accurate as Large V3?

No, and OpenAI does not claim it is. Its stated position is that Turbo “performs similarly to large-v2” across languages, with larger degradation on Thai and Cantonese specifically. Turbo also is not trained for translation, because translation data was excluded from its fine-tuning. It trades a step of accuracy for roughly 8× the decode speed.

Why does Turbo cut the decoder instead of the encoder?

Whisper’s encoder runs once per 30-second audio window, while the decoder runs once per output token. On long recordings the decoder dominates total processing time, so cutting it from 32 layers to 4 removes most of the cost. Leaving the encoder intact is what preserves the large-v3 multilingual acoustic quality.

Is a bigger Whisper model always better?

Not always. Larger variants are generally more accurate, especially on accents, background noise, and less common languages, but they use more memory and take longer on the same recording. Turbo exists precisely to make a Large-series model fast enough for a phone. For clean English audio, a smaller model often produces text you would not be able to tell apart.

Does using Whisper mean my audio goes to OpenAI?

No. Whisper is open source and can run entirely on your device, which is what LoroNote, Whisper Notes, and Aiko do. An app can also call a hosted Whisper API, in which case audio does leave your device — but that is the app’s architecture, not a property of the model.

Which Whisper app identifies speakers?

LoroNote and Whisper Notes both provide speaker labels with renaming. Aiko’s listing states that speaker detection is not currently available.

Final thoughts on choosing a Whisper app for iOS

“Powered by Whisper” has become a badge rather than a specification. Four apps can all carry it truthfully while running different model sizes, on different hardware, with your audio in completely different places.

The two questions worth asking are the ones the label does not answer: which variant, and running where. Aiko answers both in its listing and deserves credit for it. LoroNote runs the largest variant any of these puts on a phone, and runs it locally. Whisper Notes gives you three engines to switch between for the price of a coffee. Whisper Memos is a good product that happens to be a cloud service wearing the same word.

Pick by those two facts, not the badge.

Sources

• LoroNote on the App Store
• Whisper Notes on the App Store
• Aiko on the App Store
• Whisper Memos on the App Store
• OpenAI Whisper Large V3 Turbo release notes
• OpenAI Whisper model table
• whisper.cpp by Georgi Gerganov
• WhisperBoard on GitHub
• LoroNote speed figures: measured on an iPhone 15 Pro, 10 min 31 sec recording, Whisper Large V3 Turbo on device, read from the app’s in-progress display