Whisper Snapper is a native macOS app from independent developer Evan Harris that turns audio and video into text using local Parakeet, WhisperKit and Whisper models, or optional cloud APIs you supply your own keys for. It adds speaker diarization, on-device Llama 3.2 summaries and translation, and exports to TXT, SRT, VTT, Markdown, CSV and PDF.
Used this tool? Rate it
Used this tool? Rate it
Whisper Snapper is a native macOS application that converts audio and video files into written text. It is distributed through the Mac App Store under the name "Whisper Snapper Transcriptions," and the App Store record lists Evan Harris as both the seller and the developer. On Product Hunt the maker describes himself as a "solo founder. beach bum. indie-dev," which is a useful signal about what kind of product this is: a focused single-developer utility rather than a venture-funded platform with a sales team behind it.
That framing matters more than it might seem. A transcription app built by one person and sold once for a flat fee has different incentives than a subscription service that needs your recordings on its servers to justify recurring revenue. Whisper Snapper's entire architecture leans into that difference. The app is designed so that, if you want, nothing you transcribe ever touches a network connection.
The homepage states the pitch plainly: "The Transcription App Mac Users Deserve." Underneath, it explains what the app handles — "Transcribe videos, podcasts, meetings, voice memos, and more. Get AI-powered summaries and translations in 20+ languages—all running privately on your Mac."
The word "Whisper" in this app's name refers to Whisper, the open-source automatic speech recognition model family released by OpenAI. Whisper is published under a permissive open-source licence, which means any developer is free to download the model weights, run them locally, and build a product on top of them. That is exactly what Whisper Snapper does, alongside models from other sources.
Whisper Snapper is not an OpenAI product, is not operated by OpenAI, and there is no evidence of any partnership, endorsement, or corporate relationship between the two. It is an independent third-party application that runs an openly licensed model. Using the model name in a product name is legally commonplace and technically accurate here — the app genuinely runs Whisper.
It is worth stating clearly, though, that the site carries no affiliation or trademark disclaimer anywhere. The homepage names OpenAI Whisper, GPT-4o, NVIDIA Parakeet, Deepgram Nova-2, and Llama repeatedly without a line saying "not affiliated with or endorsed by these companies." That is an omission rather than a deception. Weighed against it, the app is completely open about which engine does what: the privacy policy names OpenAI and Deepgram as third-party processors and links to their policies directly. There is no attempt to pass upstream technology off as proprietary in-house invention. The honest reading is a transparent packaging of open and commercial models, with a missing boilerplate disclaimer — not a shell operation.
Most transcription products make you choose once: either you use a cloud service and accept that your recordings leave your machine, or you use a local tool and accept slower, clumsier software. Whisper Snapper's stated position is that this should be a per-file decision, not a product-level one. The site frames it as "Privacy when you need it. Speed when you don't. Run everything locally on your Mac for sensitive recordings, or use the cloud for faster processing. Whisper Snapper lets you choose—per transcription, every time."
That is the single most important thing to understand about this app. A journalist can transcribe a confidential source interview entirely offline in the morning, then push a public conference recording to a cloud API in the afternoon for speed, using the same application.
Local transcription runs entirely on your Mac's own silicon. You download a model once, and from that point the app needs no internet connection for transcription work. The App Store listing enumerates a substantial local model catalogue: "OpenAI Whisper: tiny, base, small, large-v3, large-v3-turbo, distil-large-v3" alongside "NVIDIA Parakeet: v2 (English), v3 (25 languages with diarization), TDT-CTC-110M," plus "Alibaba Qwen3 ASR" and "Llama 3.2: Local AI for translation and summarization."
The range of Whisper sizes is a genuine convenience rather than a spec-sheet flourish. Whisper tiny is fast and light and will do fine on a clean voice memo where you just need searchable text. Whisper large-v3 is far more capable on accented speech, overlapping talkers, and noisy rooms, but it is much heavier to run. Having both installed means matching the model to the job instead of always paying the worst-case compute cost.
Parakeet, developed by NVIDIA, is a different lineage of model that tends to be notably fast on Apple silicon. The v3 variant covers 25 languages and, importantly, supports speaker diarization locally — something not every offline model can do.
The cloud option is where Whisper Snapper's commercial structure differs from most competitors, and you should understand it before buying. The site says you can "use OpenAI Whisper, GPT-4o, or Deepgram Cloud APIs for speed."
Critically, the app does not resell you API access. Product Hunt's description of the app specifies transcription "in the cloud using personal API keys," and the privacy policy is unambiguous: "You must provide your own API keys for these services."
This has real consequences in both directions. On the positive side, you pay the API provider directly at their wholesale rate with no developer markup, and the app developer never becomes an intermediary holding your audio or your billing relationship. On the negative side, using cloud mode requires you to independently register with OpenAI or Deepgram, generate an API key, and fund that account. If you were expecting to pay once and have cloud transcription simply work, that is not the arrangement. Cloud usage costs are separate, ongoing, metered by the provider, and entirely outside the app's one-time price.
Diarization — labelling which speaker said which line — is what separates a usable interview transcript from a wall of undifferentiated text. The homepage describes it as: "Know who said what. Diarization available via Deepgram, GPT-4o Transcribe, and Parakeet v3."
Note the qualifier "available via." Diarization is not universal across every engine in the app. It is tied to those specific models. If speaker labels are essential to your workflow and you also require full offline operation, Parakeet v3 is the specific local model you need — the other two diarization paths are cloud services.
The July 2026 release notes for version 1.7 also record ongoing work here: "Diarization Quality Improvements - Fixed speaker segment filtering bugs that caused spurious short segments in speaker timelines." Spurious short segments are the classic diarization failure mode, where the system briefly and wrongly attributes a few words to a phantom additional speaker. That this was identified and fixed suggests active maintenance rather than a shipped-and-abandoned utility.
Version 1.1 added a second layer of AI on top of raw transcription, and unusually, this layer is also local.
Summaries run through Llama 3.2 on your own machine. The homepage describes the feature as: "Turn hour-long recordings into concise summaries with one click. Powered by Llama running entirely on your Mac—your content never leaves your device."
Translation works the same way. The homepage promises translation "to Spanish, French, German, Japanese, Chinese, and 15+ more languages" with "All processing happens locally using Llama for complete privacy." The App Store listing gives a fuller and slightly different count, advertising "AI TRANSLATION — 30 LANGUAGES" and naming Spanish, French, German, Chinese, Japanese, Korean, Portuguese, Italian, Dutch, Russian, Arabic, Hindi, Turkish, Polish, Vietnamese, Thai, Indonesian, Swedish, Danish, and Norwegian, with the note "No internet required."
This is architecturally significant. Plenty of transcription apps offer AI summaries, but nearly all of them do it by shipping your finished transcript to a cloud language model. That quietly undoes the privacy benefit of local transcription — your words still end up on someone else's server, just at a later step. Doing summarization on-device with Llama 3.2 closes that gap. For anyone handling privileged material, this is the difference between a genuinely offline pipeline and one that only looks offline.
Set expectations appropriately on quality, though. Llama 3.2 is a small model chosen so it can run comfortably on consumer hardware. Its summaries will be serviceable and its translations will convey meaning, but neither will match a frontier cloud model on nuance, idiom, or specialist terminology. That is the deliberate trade being made: privacy and zero marginal cost in exchange for some quality ceiling.
This section deserves unusual care, because accuracy is the claim buyers most want and vendors most often inflate.
Whisper Snapper is, to its credit, restrained here. Its own comparison article on choosing a transcription engine publishes no word error rates, no benchmark tables, and no percentage accuracy figures. It describes engines qualitatively — Deepgram as "Extremely fast processing," Parakeet v2 as "Fast and lightweight" — and states language coverage as "Supports 99+ languages" for the Whisper API, "Fewer languages than Whisper (30+)" for Deepgram, and "Supports 25 languages" for Parakeet v3. The homepage headline "Accurate Transcriptions on Your Mac" is marketing language with no number attached to it.
That restraint is worth more than a confident statistic would be. There is no vendor accuracy figure here to audit, because the vendor has not published one.
What can be assessed is the underlying models, since they are publicly known and independently benchmarked. Published evaluations of Whisper large-v3 report roughly 2.7% word error rate on LibriSpeech clean audio — pristine, professionally recorded, read speech. Independent testing on messier real-world material puts the range closer to 8–12% WER, and one third-party evaluation placed large-v3 at 7.88% WER. Researchers have also documented that large-v3 shows an increased tendency toward hallucination compared to earlier versions, sometimes producing long contiguous blocks of fabricated text rather than isolated word errors.
Three practical conclusions follow. First, accuracy varies enormously with recording conditions; a clean single-speaker recording and a four-person conference call over a laptop microphone are not remotely the same task. Second, accuracy varies by language — high-resource languages like English perform far better than low-resource ones, and a "99+ languages" claim describes coverage, not equal quality across all of them. Third, and most importantly for professional use, Whisper-family hallucination means errors are not always visibly garbled. The model can produce fluent, plausible, confident sentences that were never spoken. Every transcript intended for quotation, legal filing, or medical record must be checked against the audio. The app's clickable timestamps exist precisely to make that verification practical.
For a transcription tool this is the decisive section, because meeting recordings, client calls, and source interviews are among the most sensitive files most people ever create. Whisper Snapper's privacy policy is genuinely strong, and unusually specific.
The policy states that the app "does not collect, store, or transmit" personal information, usage analytics or app statistics, crash reports or diagnostics, location data, device identifiers, or "Any data for advertising purposes." Its own summary section lists "No accounts required: We don't collect personal information" and "No analytics or tracking: We don't monitor your usage."
The absence of crash reporting and analytics is a meaningful commitment, not a throwaway line. Most indie developers keep telemetry because it is how they find bugs. Forgoing it is a real cost accepted for a privacy position.
Under local transcription, the policy is categorical: "All processing occurs entirely on your Mac," "Your audio and video files are never uploaded anywhere," "No internet connection is required," and "We have no access to your files or transcripts." It labels this "the most private option available."
Under cloud transcription the policy is equally direct about the trade-off: "Your audio is sent to third-party API providers for processing." It names OpenAI (Whisper API, GPT-4o Transcribe) and Deepgram (Nova-2), links to each of their privacy policies, and adds the warning that "your audio is subject to the privacy policies of these third-party providers. We encourage you to review their policies before using cloud transcription."
That is exactly the right disclosure. It does not pretend cloud mode is private, and it points you at the parties whose terms actually govern the data.
Retention is defined by your own actions rather than a countdown clock, because there is no server holding anything. The policy lists: "Transcripts: Retained until you delete them within the app," "API Keys: Retained in Keychain until you remove them," and "Recorded Audio: Retained until you delete the files." It also notes that "Deleting the app will remove app preferences, but transcripts and Keychain items may need to be deleted separately" — a useful and often-omitted caveat.
On the two questions people most want answered — is my audio used to train AI models, and does a human ever listen to it — the policy contains no explicit statement either way. This is worth being precise about rather than generous. For local processing the question is moot: the developer states "No External Servers: We do not operate servers that receive or store your data," so there is no mechanism by which files could reach them for training or review. For cloud processing, the answer is not the developer's to give — it is governed by OpenAI's or Deepgram's terms on the account whose key you supplied. If training and human-review policy matters to you, that determination has to be made at the API provider, under the specific account tier you are using.
API key handling is specified: "If you use cloud transcription services, your API keys are stored securely in the macOS Keychain," and "We do not have access to your API keys or the audio you send." The app additionally runs sandboxed with Apple's hardened runtime enabled.
One thing does not line up, and it should be flagged. The privacy policy is stamped "Last Updated: December 2024." But the App Store record shows the app's original release date as 17 December 2025, a full year later, with version 1.7 shipping on 23 July 2026.
So the policy carries a date predating the product's public release. The most likely explanation is a simple typo — December 2024 for December 2025 — and nothing in the policy's substance contradicts how the app is described elsewhere. But taken literally, the document is stale relative to a product that has since added local summarization, translation, live transcription, and five additional models. Where a legal document and a product record disagree, the product record is the verifiable one — the App Store dates are machine-generated by Apple and cannot be edited freely, whereas the policy date is hand-typed. Treat the policy's substance as current and its date as an error, but if you are adopting this for regulated work, ask the developer to confirm and re-date it.
The homepage's audience section tells legal and medical professionals they can "Meet compliance requirements—no third-party data sharing." Read that as scoped strictly to local mode. It is accurate there — nothing is shared because nothing leaves the machine. It is not accurate in cloud mode, where audio demonstrably goes to a third party. The claim is also not a certification: there is no published HIPAA attestation, SOC 2 report, or BAA offering. For regulated work, local mode is architecturally defensible; the marketing sentence is not a substitute for your own compliance review.
The commercial model is refreshingly plain, and its headline benefit is the absence of a subscription. The homepage says you can "Try for free forever or upgrade to Pro one-time (no subscriptions ever)."
The App Store record confirms the app's download price is Free (formattedPrice "Free," price 0.00 USD). Multiple third-party listings, including Product Hunt coverage, report the Pro upgrade as a one-time purchase of $9.99 with no recurring billing.
Two honest caveats about that figure. First, the $9.99 price is not printed on the homepage itself, which states only "Pro one-time" without a number; the figure comes from third-party listings and the in-app purchase, so verify the current price in the App Store before purchasing, as developers can change it. Second, the App Store lookup record returns no itemised in-app purchase list, so the specific division of features between the free tier and Pro is not publicly documented in a way I could verify. What is free forever and what requires the upgrade should be confirmed by installing the app, which costs nothing.
The economics still deserve attention. Subscription transcription services commonly charge monthly or per-minute. A one-time payment in the region of ten dollars, combined with local models that cost nothing per hour to run, means heavy users face essentially zero marginal cost. Someone transcribing several hours weekly could recoup the price within days relative to a metered service. The counterweight, again, is that cloud mode's API charges are yours to pay separately and are not capped by the app.
This is the app's single hardest constraint, and it excludes a large share of potential users.
Whisper Snapper is macOS only. There is no Windows version, no Linux version, no iPhone or iPad app, no web interface, and no public API. The App Store record classifies it as "mac-software" with the App Store as the sole distribution channel.
The version requirement is unusually aggressive. The App Store record lists a minimum OS version of 26.0, meaning a very recent macOS release. Many Macs running perfectly serviceable older systems cannot install this app at all without upgrading the operating system first. Check your macOS version before planning around this tool — this single line disqualifies more prospective users than any other factor.
Other verified specifics from the App Store record: the download is roughly 30.4 MB (30,431,988 bytes), the app is rated 4+, it is categorised under Utilities with Business as a secondary genre, and the current version is 1.7, released 23 July 2026, against an original release of 17 December 2025.
The app's own interface language list contains only English (languageCodesISO2A: ["EN"]). Do not confuse this with transcription language coverage. The app transcribes and translates across many languages; its menus and buttons are in English only.
Supported input formats are MP4, MOV, M4A, MP3, and WAV. There is also a built-in recorder, so you can capture audio directly rather than importing it.
The July 2026 release notes describe a substantial update rather than a maintenance patch. Headlining it is "Live Transcriptions! - Record and transcribe in real time. Pause and resume recording sessions across multiple takes, all in one waveform timeline." Real-time transcription is a meaningfully different capability from batch file processing, and pause/resume across takes is a thoughtful touch for interview work.
The same release added "5 New Local Models - Cohere Transcribe (14 languages), Parakeet Japanese, Parakeet v3 INT4, SenseVoice, and an 83-language Nemotron Multilingual streaming model," rebuilt the model download system for reliability, and made performance changes including background transcript search, Cohere routing "directly to the Apple Neural Engine," and processing long audio "one chunk at a time instead of all at once."
That last change is quietly important. Chunked processing is how you stop a long recording from exhausting memory, and it suggests the developer has been fixing real problems reported by users working with genuinely long files.
A transcript that cannot leave the app is of limited use, and export coverage here is broad. The homepage lists: "Export to TXT, SRT, VTT, Markdown, CSV, or PDF. Create subtitles for video editors or formatted documents."
Each format maps to a real destination. SRT and VTT are subtitle formats that drop directly into video editors — a podcast editor cited on the homepage says "The SRT export drops right into Final Cut." Markdown suits note-taking apps and publishing pipelines. CSV opens the door to spreadsheet analysis, which matters for researchers coding qualitative interview data. PDF produces a shareable formatted document. TXT is the universal fallback.
Working with the transcript in-app is supported too. The App Store listing describes "Clickable timestamps to navigate long recordings," "Search within transcripts instantly," "Edit and refine transcripts in-app," and "All transcripts saved locally with SwiftData."
Clickable timestamps deserve emphasis given the hallucination risk discussed earlier. Being able to click a line and jump to that moment in the audio is the mechanism that makes verification fast enough to actually do. In-app editing then lets you fix what you find without exporting to another program and losing the audio linkage.
The onboarding is deliberately minimal: "Drag. Drop. Done. No accounts, no complicated settings, no tutorials needed. If you can drag a file, you can transcribe it." The absence of any account requirement is consistent with the no-server architecture — there is nothing to create an account on.
Bulk processing is supported, with multiple files added at once and per-file progress tracking, which matters if you are working through a backlog of recordings rather than one file at a time.
The homepage identifies six audiences, and they are worth evaluating critically rather than accepting wholesale.
Journalists and researchers are the strongest fit. The stated benefit — transcribe interviews "without uploading sensitive sources to the cloud" — addresses a genuine professional obligation. Source protection is not a preference for reporters; it can be an ethical and sometimes legal duty. Local-only processing with no server component is a real answer to that, not a marketing gesture.
Legal and medical professionals are the second strongest fit, with the caveats already stated: local mode only, and the marketing compliance line is not a certification.
Podcasters and content creators benefit primarily from diarization plus SRT export. The workflow of transcribing a multi-host episode, getting speaker labels automatically, and dropping subtitles into an editor is exactly what the tool is shaped for.
Students and educators get lecture transcripts and local summarization, though this group should weigh the macOS 26 requirement carefully, since student hardware is often older.
Business and remote teams can capture meeting notes, but note that this is a single-user desktop app. There is no shared workspace, no team billing, no collaborative editing, and no meeting-bot that joins calls automatically. Sharing means exporting a file and sending it. Teams wanting integrated collaboration should look elsewhere.
Accessibility use cases — producing captions for hearing-impaired audiences — are well served by the subtitle export formats, with the standard caveat that machine captions require human review before publication.
Who should not choose this: anyone not on a recent Mac, anyone needing mobile capture, anyone needing an API for automated pipelines, anyone requiring team collaboration features, and anyone who wants cloud transcription to work without setting up their own API accounts.
Collected in one place, so nothing is buried:
Platform lock-in. macOS only, minimum OS 26.0. No Windows, Linux, iOS, Android, or web version. This is the biggest single limitation.
No API. Nothing here can be automated into a server-side workflow. It is an interactive desktop application.
Cloud mode is bring-your-own-key. Requires separate OpenAI or Deepgram accounts, and the usage costs are yours, uncapped by the app.
Diarization is model-dependent. Only Deepgram, GPT-4o Transcribe, and Parakeet v3 provide it. Offline diarization means Parakeet v3 specifically.
Local model performance depends on your hardware. Larger Whisper models are demanding. On an older or lower-specification Mac, large-v3 will be slow. The app cannot make your machine faster; it can only let you pick a smaller model.
Storage cost. Local models must be downloaded and stored. The 30.4 MB app size is the installer, not the models — a full local model collection consumes considerably more disk.
Summarization and translation quality is bounded by a small model. Llama 3.2 running on-device will not match frontier cloud models.
No independent review corpus. The App Store record shows a user rating count of 0 and no average rating. There is no G2, Capterra, or Trustpilot presence, and no coverage from major technology press. Product Hunt provides 139 upvotes and a #7 daily rank, which is a real but modest signal. The homepage testimonials are first-name-and-role attributions that cannot be independently verified. This is normal for a young indie app, but it means you are buying largely on the strength of the free tier rather than an established review consensus.
Interface is English-only, regardless of transcription language coverage.
Accuracy is inherently variable, and Whisper-family hallucination means errors can look fluent. Always verify quotable material.
The privacy policy date is inconsistent with the product timeline, as documented above.
A few practical habits materially improve output.
Start with the free tier before paying anything, and test it on your own worst-case audio rather than a clean sample. The recording you actually struggle with is the one that determines whether the tool is useful to you.
Match the model to the material. Use a small local model for clean single-speaker audio where speed matters; reserve large-v3 for difficult recordings with accents, crosstalk, or background noise.
Decide privacy per file, deliberately. The whole point of the architecture is that this is a per-transcription choice. Make it consciously rather than defaulting.
Improve the input before blaming the model. An external microphone, a quieter room, and speakers who avoid talking over each other will do more for accuracy than any model upgrade.
For interviews, enable diarization from the start using a supporting engine. Retro-fitting speaker labels by hand is far slower than getting them automatically.
Verify before you quote. Use the clickable timestamps to spot-check any passage you intend to publish, cite, or file.
If you use cloud mode, set spending limits at the API provider. The app does not cap your usage, and metered transcription of a large backlog can accumulate cost quickly.
No. Whisper Snapper is an independent macOS app whose App Store record lists Evan Harris as the developer and seller. "Whisper" in the name refers to OpenAI's openly licensed speech recognition model, which the app runs alongside models from NVIDIA, Deepgram, Cohere, Alibaba and Meta. There is no evidence of any partnership or endorsement. The site carries no explicit affiliation disclaimer, which is an omission worth noting, but the app is otherwise transparent about which third-party engines it uses and links to their policies directly.
Only if you choose a cloud engine. With local models the privacy policy states plainly that "Your audio and video files are never uploaded anywhere" and no internet connection is required. If you select OpenAI or Deepgram cloud transcription, "Your audio is sent to third-party API providers for processing." The choice is made per transcription, so you control it file by file.
There is no server-side retention period, because the developer states they operate no servers that receive your data. Everything is kept on your own machine until you delete it: transcripts "Retained until you delete them within the app," recorded audio "Retained until you delete the files," and API keys in the Keychain until you remove them. Note that uninstalling the app does not automatically remove transcripts or Keychain entries.
The privacy policy makes no explicit statement on either point. For local processing this is moot — the developer operates no servers, so no mechanism exists for your files to reach them. For cloud processing, the answer is set by OpenAI's or Deepgram's terms on the account whose API key you supplied, not by this app. If training and human-review policy is important to you, verify it at the API provider under your specific account tier.
The app downloads free from the Mac App Store, and the developer's position is "Try for free forever or upgrade to Pro one-time (no subscriptions ever)." Third-party listings report the Pro upgrade as a one-time $9.99 purchase. That figure is not printed on the homepage, so confirm the current price in the App Store. Cloud API usage, if you enable it, is billed separately by OpenAI or Deepgram and is not included.
Only if it is a Mac running macOS 26.0 or later — that is the minimum OS version in the App Store record. There is no Windows, Linux, iOS, Android or web version, and no API. This requirement rules out a significant number of otherwise capable machines, so check your macOS version first.
The developer publishes no accuracy percentage, which is more honest than inventing one. Accuracy depends on the model you pick, your audio quality, and the language. Independent benchmarks of Whisper large-v3 report roughly 2.7% word error rate on pristine read speech but around 8–12% on realistic recordings. Researchers have also documented that this model family can hallucinate fluent text that was never spoken, so any passage you intend to quote or file should be checked against the audio.
Yes, but only with specific engines. Diarization is "available via Deepgram, GPT-4o Transcribe, and Parakeet v3." If you need speaker labels while staying fully offline, Parakeet v3 is the model to use, since the other two options are cloud services. Version 1.7 also fixed bugs that produced spurious short speaker segments.
Export covers TXT, SRT, VTT, Markdown, CSV, and PDF, which handles subtitles for video editors, notes, spreadsheets, and shareable documents. Before exporting you can edit transcripts in-app, search them instantly, and use clickable timestamps to jump to the matching audio — which is the practical way to verify anything questionable.
Yes, and this is unusual. Both run on Llama 3.2 on your device, with the App Store listing specifying "No internet required" for translation across its language list. Most competing apps send transcripts to a cloud model for summarization, which quietly undermines local transcription's privacy benefit. The trade-off is that a small on-device model will not match a frontier cloud model on nuance or specialist terminology.
