@iPhone_News
Aqua Voice shows just how good Mac dictation could be if Apple really tried
Shared desktop and iPhone support, plus Android on Aqua. Aqua Voice runs its own model, Avalon, with accuracy you can verify, at $8/mo against Willow’s $12 on the same annual billing. Here’s the honest comparison.
With Realtime Mode on Max, Aqua Voice refines your words as you speak: punctuation, formatting, and technical terms land in real time.
Both are Y Combinator startups shipping real-time dictation on Mac, Windows, and iPhone, and Aqua Voice now adds Android, where Willow still lists it as coming soon. Both are HIPAA compliant, too. Aqua Voice has been at this since YC’s W24 batch in early 2024; Willow is a Spring 2025 company. Willow bets on breadth: 100+ languages and a free tier with no word cap. Aqua Voice bets on the model itself, one it owns and benchmarks publicly, at a lower price, with a developer API behind it.
You want accuracy you can verify against a public benchmark and an independent press test; you would rather pay $8/month than $12; or you dictate code and technical terms and want a model purpose-built for them, with a developer API behind it.
You dictate in a language outside the 49 Aqua Voice covers, or you want to keep using a dictation app for free rather than upgrade.
Both companies publish accuracy claims, measured in different ways. Avalon, the model inside Aqua Voice, scores 97.4% on AISpeak-10, our public benchmark of AI and technical vocabulary, and ranked #1 among proprietary systems on the OpenASR leaderboard, an independent benchmark maintained by Hugging Face, at its October 2025 debut. Independent reviewers have checked it too: 9to5Mac measured Aqua at 1 error to Apple Dictation’s 17 on the same passage.
On AISpeak-10, the published key-term subset, Avalon scores 97.4% against ElevenLabs Scribe at 78.8% and Whisper Large v3 at 65.1%. Willow hasn’t disclosed which speech model it runs. It publishes a zero-edit-rate benchmark from its own testing, with no word error rate, and Aqua was not in that test, so try both on your own speech. You can see Avalon’s results on the OpenASR leaderboard and read more in our Avalon write-up.
AISpeak-10 is Aqua’s English technical-speech dataset. Avalon transcribed the key term correctly in 97.4% of clips, versus 65.1% for Whisper Large v3. This measures key-term recognition, not overall transcript accuracy or a head-to-head test of these apps. See the evaluation and results.
The head-to-head on model, accuracy, languages, privacy, and price. Where a number is Willow’s own claim rather than a verified figure, it’s marked.
| Feature | Aqua Voice | Willow Voice |
|---|---|---|
| Speech model | Avalon (own, benchmarked publicly) | Undisclosed (Frontier Mini / Pro) |
| Public benchmark | #1 proprietary on OpenASR (Oct 2025 debut) | Own zero-edit-rate test (Oct 2026) |
| Technical-term accuracy | 97.4% (AISpeak-10, our benchmark) | Not published |
| Independent press test | ✅ 9to5Mac (1 error vs Apple's 17) | None published |
| Latency (both vendor-stated) | ~450ms after you stop | "as little as 200ms" |
| Mac / Windows / iPhone | ✅ | ✅ |
| Android | ✅ | ❌ ("coming soon") |
| Languages | 49, auto-detected | 100+ claimed, no list published |
| Free tier | 1,000 lifetime words | Unlimited on its weaker model |
| Custom dictionary | ✅ up to 800 terms on Pro | ✅ |
| Reads the active app | ✅ | ✅ |
| Offline mode | ❌ cloud-based | Advertised, undocumented |
| SOC 2 Type II | ✅ | ✅ |
| ISO 27001 | ✅ certified | Not stated |
| Zero Data Retention | ✅ (Business) | ✅ (Business, per Willow) |
| GDPR | ✅ compliant | Not stated |
| HIPAA BAA | ✅ sign in the app | ✅ sign in the app |
| Developer API | Avalon API ($0.39/hr) | None public |
| Pro price | $8/mo annually, $10 monthly | $12/mo annually, $15 monthly |
| Teams | $19 per seat/mo annually, 2 seats minimum | $28/mo annually, 3 seats minimum |
| Best for | Verifiable accuracy, price, developers, Android | Language breadth, staying free |
Speech model
Public benchmark
Technical-term accuracy
Independent press test
Latency (both vendor-stated)
Mac / Windows / iPhone
Android
Languages
Free tier
Custom dictionary
Reads the active app
Offline mode
SOC 2 Type II
ISO 27001
Zero Data Retention
GDPR
HIPAA BAA
Developer API
Pro price
Teams
Best for
Aqua Voice is SOC 2 Type II and ISO 27001 certified and GDPR and HIPAA compliant. Willow says it is SOC 2 Type II and HIPAA compliant too. If you want the model rather than the app, Aqua exposes Avalon through the Avalon API. And if Wispr Flow is also on your shortlist, there is a side-by-side for that too.
Avalon's results are public: 97.4% on AISpeak-10, Aqua's benchmark of AI and technical speech, #1 among proprietary systems on the independent OpenASR leaderboard at its October 2025 debut, and a 9to5Mac test measuring 1 error against Apple Dictation's 17. Willow's accuracy figures come from its own tests, with no third-party result to check.
Aqua trains and operates Avalon itself, which is why it can publish benchmarks and keep tuning the model for the speech its users actually produce. Willow brands its tiers Frontier Mini and Frontier Pro but publishes nothing about what runs behind those names, and its security page refers only to unnamed AI subprocessors. This is the most defensible difference between the two: one of these companies has a third-party benchmark result to point to and the other does not.
The Avalon API exposes the model that runs the app, OpenAI-compatible, at $0.39 per hour of audio. If you want dictation inside your own product rather than only in an app, Aqua is something you can build on and Willow lists no public developer API.
Most speech models are trained on audiobook and meeting speech, so they mangle kubectl, tsconfig, zshrc and the names of the tools you actually work in. Avalon was trained on human-computer interaction speech instead: code, prompts and terminal commands. The desktop app reads the active window so it writes casually in Slack and code-aware in Cursor or Claude Code, and Pro carries a custom dictionary of up to 800 terms for the names no model can know. Willow publishes no separate score for technical vocabulary, the words that break most dictation apps.
Aqua Voice Pro is $8/month billed annually ($96/year), or $10 month to month. Willow Pro is $12/month billed annually, or $15 month to month. Compare annual against annual or monthly against monthly and Aqua costs about a third less either way, and the Free plan lets you test the accuracy claim before paying anything.
Aqua Voice runs on Mac, Windows, iPhone, and Android. Willow lists Android as coming soon, so if you need dictation on an Android phone today, Aqua Voice is the one that ships it.
Willow advertises 100+ languages, against Aqua Voice's 49 with automatic detection. Willow publishes no language list and its own help center still says 50, so the exact number is unauditable, but breadth is clearly their bet and not ours. If you dictate in a language Aqua doesn't cover, check Willow.
Willow's free plan offers unlimited dictation on its weaker Frontier Mini model. Aqua Voice's Free plan is 1,000 lifetime words, roughly seven minutes of talking, and then you decide. If you want to use a dictation app indefinitely without paying, Willow lets you and Aqua Voice doesn't.
Used by people at
Students get 70% off Pro and Max.
Yes, and Aqua Voice now covers one more platform: it runs on Mac, Windows, iPhone, and Android, while Willow still lists Android as coming soon. Aqua Voice runs its own model, Avalon, with published benchmarks and an independent 9to5Mac press test, and Pro costs $8/month billed annually against Willow Pro at $12. Willow advertises 100+ languages to Aqua's 49 and a free tier with no word cap. If accuracy you can verify and price matter most, Aqua Voice is the stronger pick.
It depends on what you weigh most, and this page lays out both sides. Aqua Voice is the stronger pick if you want accuracy you can verify (97.4% on the AISpeak-10 benchmark, plus #1 among proprietary systems on the OpenASR leaderboard at its October 2025 debut), a lower price, or a developer API. What separates them is the speech model: Aqua owns and benchmarks its own, and Willow has not disclosed what powers its product. Willow is the better pick if you need a language outside Aqua's 49 or a free tier without a word cap.
Both are dictation apps that clean up speech and type into any app on Mac, Windows, and iPhone, in real time on the desktop, and Aqua Voice also runs on Android. The core difference is the speech engine. Aqua Voice runs Avalon, a model it trained itself and benchmarks publicly, and is tuned for code, AI prompts, and technical vocabulary. Willow names its model tiers Frontier Mini and Frontier Pro but does not disclose what powers them, and its accuracy figures come from its own tests. Aqua Pro is $8/month billed annually; Willow Pro is $12.
Both publish accuracy claims, measured in different ways. Avalon scores 97.4% on AISpeak-10, Aqua's public benchmark of AI and technical speech, and ranked #1 among proprietary systems on the independent OpenASR leaderboard at its October 2025 debut. 9to5Mac tested Aqua against Apple Dictation and measured 1 error to Apple's 17 on the same passage. Willow publishes a zero-edit-rate benchmark from its own testing (95% in English) and claims 98%+ accuracy, with no word error rate or third-party result. Aqua was not in Willow's test, so try both on your own speech.
Aqua Voice runs Avalon, a model it built and benchmarks in public: 97.4% on AISpeak-10, Aqua's benchmark of AI and technical terms, and #1 among proprietary systems on the OpenASR leaderboard at its October 2025 debut. Willow has not publicly disclosed what powers its dictation. It brands its model tiers Frontier Mini (free) and Frontier Pro (paid), but nothing is published about what runs behind those names, and its security page refers only to unnamed AI subprocessors. If knowing what transcribes your voice matters to you, that is the clearest difference between the two.
Yes, on both billing periods. Aqua Voice Pro is $8/month billed annually ($96/year), or $10 month to month. Willow Pro is $12/month billed annually, or $15 month to month. Willow does have the more forgiving free tier: it offers unlimited dictation on its weaker Frontier Mini model, against Aqua Voice's 1,000 lifetime words on the Free plan. So the cheaper paid plan and the more generous free plan are on opposite sides.
Yes. Aqua Voice runs on Mac, Windows, iPhone, and Android. Willow runs on Mac, Windows, and iPhone and lists Android as coming soon, so the two match everywhere except Android, where only Aqua Voice ships today.
Yes. Aqua Voice launched on iOS in April 2026 as an AI voice keyboard, so you can dictate into any app on your phone and correct what you said by voice. Willow also runs on iPhone, so this one is a match. A cross-platform Pro subscription covers Mac, Windows, iPhone, and Android together rather than being billed per device.
Yes. Aqua Voice has a SOC 2 Type II audit report and is ISO 27001 certified and independently audited, and it is GDPR and HIPAA compliant, with a public trust center linked from the site footer. Every plan includes a Privacy Mode, and Business lets you apply it across your team. Willow also says it is SOC 2 Type II and HIPAA compliant, but its security page does not mention ISO 27001 or GDPR.
Yes. Business includes everything in Max, plus centralized billing, single sign-on, automatic member provisioning, Zero Data Retention, reporting, and shared dictionaries. It costs $19 per seat per month billed annually or $24 per seat billed monthly, with a minimum of 2 seats. Contact sales for 100 or more seats.
Yes. The Avalon API is an OpenAI-compatible speech-to-text endpoint at $0.39 per hour of audio, running the same model as the app, so you can put dictation inside your own product rather than only use it in an app. Willow lists no public transcription API, only an enterprise usage-reporting API.
Yes, it's what Aqua Voice is built for. Avalon was trained on the speech developers and AI users actually produce, reaching 97.4% accuracy on technical terms like programming keywords, framework names, and CLI commands such as kubectl. The desktop app also reads the active application to adjust tone, so it writes differently in Slack than it does in Cursor or Claude Code, and Pro carries a custom dictionary of up to 800 terms for the names it cannot know.