Oh So Video
FreeFree captions generator that runs Whisper in your browser, so the video never leaves your device
⚡ Quick Verdict
Anyone captioning video they cannot or would rather not upload, on a computer or a phone, and creators who want burned-in captions with no account or watermark
Chinese, Japanese or Thai speech, translation, speaker labels, batch work, or buyers who need terms of service and a named supplier
Free. No paid tier, no account and no watermark
Everything is free; the vendor states there is no premium tier
Transcription runs on your own device, and the site's code confirms the video is not uploaded
Smaller Whisper models on the CPU only: slow on long videos, and no Chinese, Japanese or Thai
Bottom line: Oh So Video scores 4.1/5. It is a free captions generator that keeps the video on your device, with no account, no watermark and four export formats, and its privacy claims hold up in the site's code. On a 37-minute NASA podcast its captions came out with a 4.6% word error rate. It is held back by smaller speech models than MacWhisper Pro offers, CPU-only processing that is slow on long videos, no translation or speaker labels, seven missing languages including Chinese and Japanese, and a vendor too new to have terms of service or a named operator.
What Oh So Video is
Oh So Video is a free captions generator that runs in your web browser. You drop in a video, it transcribes the speech on your own computer or phone, you correct any lines it got wrong, and you export the captions burned into an MP4 or as an SRT, VTT or plain text file. The video is not uploaded: the speech model is downloaded to your browser and does the work there.
The captions generator is one of 15 free tools on ohsovideo.com, alongside tools that compress, trim, crop and convert video, blur faces, and a separate transcript generator that uses the same speech engine. All of them follow one idea: the page brings the software to your device instead of sending your file to a server. There is no account, no paid tier and no watermark.
The site was built by Tom Limb, who previously founded the stock footage site Videvo, sold to Freepik in 2022 (Freepik now operates as Magnific). The domain was registered in September 2026, so the product is weeks old rather than years old, and the site names Tom Limb as founder but no company that operates it.
This review assesses Oh So Video against our published 8-parameter framework. A browser tool ships its working code to every visitor, so much of what follows was confirmed in the site's code, and we measured accuracy on a full 37-minute podcast episode. Where a point is the vendor's word rather than something the code or a live check shows, the page says so.
How it works
When you choose a file, the page reads it from your device and decodes the sound to 16 kHz mono audio inside the browser. A voice activity detector, Silero VAD, finds the stretches that contain speech, so silent parts are skipped rather than transcribed. The audio then goes to a background worker that runs OpenAI's open-source Whisper speech model through transformers.js, Hugging Face's JavaScript library for running models, on ONNX Runtime Web. Whisper writes a time for every word, which is what lets the captions appear as each word is spoken.
The model you get depends on the device. On a computer the page loads Whisper small, a model with 244 million parameters, quantized to 8 bits; for English it uses the English-only version of small. On phones and tablets, and on any device that reports 4 GB of memory or less, it uses Whisper base, with 74 million parameters, also at 8 bits. When the language is left on automatic, the multilingual base model listens first, and on a computer a result that is not clearly English gets a second opinion from multilingual small. If a computer runs out of memory, the page drops back to base. The tiny, medium, large and turbo sizes are never used.
Transcription runs on the processor (CPU) through WebAssembly. The code contains a path for the graphics chip (WebGPU), but only a debug setting in the page address switches it on, so ordinary visitors do not get it. That is the main reason long videos take time: the vendor's own estimate is about 10 minutes of work per hour of video on a computer, and longer on a phone. In our test, transcribing a 37-minute episode took 11.4 minutes on an M1 Pro laptop, closer to 18 minutes per hour.
The model files come from the vendor's own content delivery network, cdn.ohsovideo.com, rather than straight from Hugging Face. They are copies of the public Whisper conversions published by the onnx-community group on Hugging Face: the encoder and decoder files have identical sizes, and the file we hashed has an identical SHA-256 checksum. With automatic language detection, the first visit downloads about 340 MB on a computer, because both the base detector and the small model are fetched, and about 155 to 170 MB on a phone. The browser then keeps the files, which is why the site says the tool works offline once loaded.
Burned-in captions are drawn onto each frame in the browser and encoded as H.264 video with AAC sound by WebCodecs, the browser's built-in video encoder, through the open-source Mediabunny library. No step needs a server, and we found no upload endpoint anywhere in the site's code.
Accuracy on real speech
Accuracy decides whether a captions tool saves time, and it is the one question the code cannot settle. What the code does settle is the ceiling. Whisper small and base are the smaller sizes in the Whisper family, both quantized to 8 bits, and phones, tablets and low-memory devices get base. For comparison, MacWhisper's free tier runs the same small and base sizes, while its Pro license adds Large v3, Turbo and Medium. The vendor itself says accuracy is better in widely spoken languages, and in English in particular, and lower on a phone.
Every line is editable before export, and the vendor says corrections keep their timing. The one we made did: we typed over a misheard name, and the re-exported subtitle file changed in that one line, with every timestamp the same. The site also says the editor flags words that are often misheard in English; no flag appeared in our run.
To measure accuracy, we captioned a full episode of NASA's Houston We Have a Podcast (episode 442, 37 minutes): a host recorded in a studio, four crew members recorded inside the CHAPEA Mars analog habitat, and an interview with the study's principal investigator. We ran it on a laptop with the language left on automatic, which chose English and the English-only small model, and compared every word with the transcript NASA publishes for the episode.
The word error rate came out at 4.6%: 2.2% for the studio host, 2.9% for the interview and 6.8% for the habitat recordings. That figure leaves out two short montages of archival audio clips that NASA's transcript does not include; counting them as errors, it is 5.6%. We also forced the page onto its phone setting on the same laptop, where it switches to the smaller Whisper base model: the same episode came out at 7.4%, and twice the page caught the model repeating itself and cut the loop.
What went wrong is what usually goes wrong with Whisper. The mission acronym came out wrong 8 of the 24 times it was said, spelled four different ways, and the host, Kenna Pell, came out as Ken Appel; the other names came out right. Once, mid-interview, the model filled a one-second pause with a sentence nobody said, 14 words about the pandemic, on screen for under two seconds. Timing did not drift: at the six clear pauses we could measure, between the 21st and 35th minutes, the next caption started within 0.4 seconds of the speech resuming. Read the captions through before publishing, and check names and acronyms.
Oh So Video pricing
There is nothing to pay. Oh So Video has no paid tier, no trial, no account and no watermark on exports, and the vendor states there is no premium tier to upgrade to and no limit on how many videos you caption. We found no sign-in, upsell or advertising code in the site's scripts, which fits that description.
- Price: $0, for every feature on the page.
- Account: none exists, so there is nothing to sign up for.
- Watermark: none on exported video, and none in the MP4 from our test.
- Limits: per video only, 3 hours on a computer and 1 hour on a phone, with no file-size limit.
On how it stays free, the vendor says the tools cost almost nothing to run because the work happens on your device rather than on its servers, and that there is no investor money behind it, nothing harvested and no ads. The costs that remain are hosting and the one-time model download each new device makes. No business model is stated beyond that. It is not a defect today, but it is worth knowing before you build a routine around a free tool from a new vendor.
Languages, limits and exports
Languages
The language menu lists 92 languages plus automatic detection, which the vendor says works from the first half-minute of speech. That is the full Whisper list for these models minus seven languages that are blocked in the code: Chinese, Japanese, Thai, Lao, Burmese, Khmer and Tibetan. The vendor says they are left out because they are written without spaces between words. If automatic detection lands on one of them, the page says the language is not supported yet. The captions are always in the language spoken: the tool transcribes and never translates.
Limits
- Length: up to 3 hours per video on a computer and 1 hour on a phone. The phone cap applies to every device the page treats as a phone: touch-first devices such as tablets, and any device that reports 4 GB of memory or less. Both caps are in the code, and the vendor confirms them.
- File size: no limit. The captions code has no size check, and the vendor states there is none. Files over 2 GB are not kept for resuming after a reload.
- Input: MP4, MOV, WebM and MKV.
- First use: about 340 MB of model files on a computer and about 155 to 170 MB on a phone. The vendor suggests WiFi on a phone.
- Speed: about 10 minutes of work per hour of video on a computer by the vendor's estimate, and longer on a phone. Transcribing our 37-minute test episode took 11.4 minutes on an M1 Pro laptop, and 12.4 from choosing the file, with the first-visit model download. On a phone, keep the page open: switching apps pauses it.
Exports
- MP4 with burned-in captions in five styles (Classic, Minimal, Bold, Highlight and Pill), with an optional animation (Bounce, Pop, Fade or Reveal), a size, a position and a highlight color. If your browser cannot encode H.264 video, the page says so and points you to the SRT, or to making the MP4 on a computer or in another browser.
- SRT and VTT subtitle files with the same text and timings, for YouTube, video editors and web players.
- TXT, the words alone, one line per caption line, with no times.
- Pacing presets run from one-to-three-word captions for short-form video to two-line subtitles. The code sets one-line captions at up to 42 characters and 4 seconds, and the classic two-line preset at up to 42 characters per row and 7 seconds.
Privacy: what stays on your device, and what is sent
The core claim holds up. In the captions tool the video is read, decoded and transcribed inside the browser, and we found no code path that sends the video, its audio, a frame, the transcript or the file name anywhere. The only outgoing data requests in the captions code go to PostHog's EU cloud, and they carry technical reports, not media.
What is sent automatically
When a tool fails, or stops partway because the page was closed, the page sends a short error report to PostHog's EU cloud without asking. It contains the tool and page, the error code and message, the file's type and extension and a rough size band (for example 10 to 100 MB), the codec and picture size and how far an export got when an export fails, and your browser, operating system and whether you are on a phone, tablet or computer. It never includes the video, its name or its contents. Each report carries a random ID created for that report rather than a lasting identifier, the request asks PostHog not to look up your location or build a person profile, and the privacy policy says IP addresses are not stored.
There is no opt-in and no switch on the page to turn these reports off. The vendor's education page says the PostHog address can be blocked by a network filter without breaking the tools. Separately, visits are counted by Plausible Analytics and Cloudflare Web Analytics without cookies, and tool steps are counted with fixed event names such as "Captions export .srt". Two cards send data only when you submit them: Report a bug, which includes browser and file-type details unless you untick a box, and an optional feedback card, shown once after you have used a tool twice, which adds technical details only if you tick a box.
What is kept on your device
So that a reload, or a phone closing the tab, does not lose your work, the page keeps the video and the lines written so far in the browser's own storage (IndexedDB) while captions are being made, for up to a day. The transcripts of your last five videos, with your edits, are kept for 14 days. None of it leaves the device, and the Clear saved work button deletes it. On a shared or classroom computer, that local copy is the thing to remember.
The Google logger on the Blur faces tool, now blocked
One discrepancy turned up outside the captions tool. The site's Blur faces tool uses Google's open-source MediaPipe library, which includes a usage logger coded to post task and timing data, not video, to a Google address, odml.pa.googleapis.com. On 25 October 2026 that logger was bundled with nothing to stop it, while the privacy policy said nothing loads from Google. Whether it sent anything in practice before the change is not established.
As of 3 November 2026, the Blur faces worker replaces the browser's fetch function as it loads, so any request to that address is answered inside the browser with an empty 204 response and never leaves the device. The logger code is still bundled, but its requests are blocked. We confirmed this in the site's code rather than by capturing network traffic. The privacy policy, dated 17 October 2026, says nothing loads from Google, which now matches. The captions generator does not load MediaPipe at all.
What we verified, and what we did not
This is where the line between confirmed and vendor-stated is drawn, so you can weight the rest of the page yourself.
Confirmed in the site's code and on its live pages
- Transcription runs in a background worker in the browser, with transformers.js on ONNX Runtime Web, on the CPU unless a debug setting is used.
- The model choice: Whisper small on computers, English-only small for English, and Whisper base on phones, tablets and devices reporting 4 GB of memory or less, all 8-bit.
- The model files load from cdn.ohsovideo.com and match the onnx-community files on Hugging Face.
- The language menu has 92 languages, the seven exclusions are hard-coded, and the task is always transcription, never translation.
- The length caps of 3 hours and 1 hour, and the absence of a file-size check in the captions tool.
- The four exports: MP4 with burned-in captions, SRT, VTT and TXT, all four produced in our test, with no watermark in the frames of the MP4 we checked.
- No sign-in, upsell, advertising or watermark-drawing code in the site's scripts.
- The only outgoing POST requests in the captions code go to PostHog's EU cloud, and none of them carries the video, audio, frames, transcript text or file name.
- The Blur faces worker blocks its bundled Google logger inside the browser, as of 3 November 2026.
- The site has no terms of service page: /terms and /terms-of-service both return a 404.
- Accuracy on a 37-minute NASA podcast: a 4.6% word error rate against NASA's transcript (5.6% counting two archival clips it leaves out), with no timing drift at the pauses we measured.
- A correction typed into a line kept its timing in the exported subtitle file, in the one edit we made.
Vendor-stated, not independently confirmed
- That the editor flags often-misheard English words for checking. No flag appeared in our run.
- That the tool works offline once the page and the model have loaded. The model files are cached by the browser, which is consistent with it.
- The reason for the seven missing languages: that they are written without spaces between words.
- That accuracy is better in widely spoken languages and in English, and lower on a phone.
- That there is no limit on the number of videos, and no premium tier planned.
- That PostHog and Plausible do not store IP addresses.
Not documented anywhere we could find
- Terms of service. No terms page exists or is linked.
- The legal entity behind the service. The site names Tom Limb as founder, but no company, company number or address.
- A business model beyond the vendor's statement that the tools are cheap to run.
- Independent reviews or benchmarks. Beyond launch and directory listings, we found none.
The site's claims, side by side
Oh So Video makes specific claims, which is what made them checkable. The table sets each one against what we could establish.
| What Oh So Video says | What we could establish |
|---|---|
| Your video never leaves your device | Confirmed in the site's code for the captions tool. The video, audio, frames, transcript and file name are not sent anywhere. Error reports, visit counts and optional feedback are sent, and none contains media |
| Captions in more than 90 languages | Confirmed: 92 in the language menu. Chinese, Japanese, Thai, Lao, Burmese, Khmer and Tibetan are blocked in the code. Transcription only, no translation |
| Free, no watermark, no account | Consistent with the code: no sign-in, upsell or advertising code, and no code that draws a watermark. The vendor states there is no premium tier |
| Any size, up to 3 hours (1 hour on a phone) | Confirmed in the code: caps of 10,800 and 3,600 seconds and no file-size check. Tablets and devices reporting 4 GB of memory or less get the phone cap |
| About 340 MB on a computer, up to 170 MB on a phone | Consistent with the model file sizes on the vendor's CDN |
| Works offline once loaded | Vendor-stated. The model files stay in the browser's cache after the first use, which offline use requires |
| About 10 minutes of work per hour of video on a computer | Slower in our test: transcribing a 37-minute episode took 11.4 minutes on an M1 Pro laptop in Chrome, about 18 minutes per hour. It depends on the device, since transcription runs on the CPU |
| Clear speech in a quiet room transcribes very well | Holds on our test: 2.2% word error rate for a studio host and 2.9% for an interview on a NASA podcast, and 6.8% for crew recorded inside a habitat |
| Nothing loads from Google (privacy policy) | True as of 3 November 2026: the Blur faces tool now blocks its bundled Google logger inside the browser. On 25 October the logger was bundled without that block |
Two things stand out. The vendor documents its own data flows in unusual detail, down to a list of every host a tool page contacts, and that documentation matches the code with one exception, the Blur faces logger, which is now blocked in the code. What the vendor does not document is the paperwork: no terms of service and no named operator.
How Oh So Video compares
Other tools' details below are as shown on their ToolChase reviews and in our AI caption generator guide.
| Tool | Where the audio is transcribed | Free tier | Caption languages | Caption output | Runs on | Score |
|---|---|---|---|---|---|---|
| Oh So Video | On your device, in the browser | Everything free, no account or watermark | 92 | Burned-in MP4, SRT, VTT, TXT | Current browsers on computers and phones | 4.1 |
| MacWhisper | On your Mac | Free tier with smaller models; Pro €65 one-time | Not stated | SRT, VTT, TXT, CSV, DOCX | macOS only | 4.5 |
| Clipchamp | Microsoft's servers | Free, 1080p, no watermark, Microsoft account | Not stated | Burn-in | Browser, and Windows 11 | 4.1 |
| CapCut | In the cloud | Free, watermarks on some premium assets | 50+ | Burn-in, SRT | Desktop, mobile and web | 4.4 |
| VEED | In the cloud | Free tier with watermarked exports | 100+ | Burn-in, SRT | Browser | 4.4 |
| HappyScribe | In the cloud | 10-minute one-time trial | 150+ (AI) | SRT, VTT, burn-in | Browser, plus mobile apps | 4.3 |
"Not stated" means our review of that tool does not publish a figure, so we do not guess one.
Against MacWhisper
MacWhisper is the one tool in our caption guide that, like Oh So Video, keeps your audio on your own machine, so the comparison is close in spirit. Both run Whisper locally. MacWhisper is a native Mac app: its free tier runs Whisper small and base, the same sizes Oh So Video uses, and its Pro license, €65 one-time, adds Large v3, Turbo and Medium, batch transcription and speaker diarization. It exports TXT, SRT, VTT, CSV and DOCX, and our guide notes it produces subtitle files rather than styled captions.
Oh So Video runs in a browser on any current computer or phone with nothing to install, and adds styled burned-in captions. Choose MacWhisper on a Mac when difficult audio, several speakers or batch work make the larger models worth paying for. Choose Oh So Video on any other device, or when you want a finished captioned MP4.
Against the cloud tools
As our caption guide notes, every other tool in it processes audio in the cloud, which means uploading it. In exchange they offer things Oh So Video does not: translation (VEED across 100+ languages, Kapwing across 60+, HappyScribe between dozens), speaker labels (HappyScribe, and speaker identification in CapCut), optional human transcription (HappyScribe, from $2.00 per minute), and full video editors (CapCut, Clipchamp, Kapwing, VEED). Their free tiers come with a watermark, an account or a trial.
The trade is simple to state. If the footage may be uploaded and you need translation, speaker labels or an editor, a cloud tool does more. If it may not, or you only need captions, Oh So Video does the job without the upload, the account or the watermark.
What Oh So Video does well
1. The privacy claim survives a reading of the code
Plenty of tools say your files never leave your device. Here the claim could be checked, because a browser tool ships its code to every visitor. The captions code has no upload path, and its only outgoing requests carry error and usage reports without media. For interviews, client work, school video or anything you are not allowed to upload, that is the reason to use it.
2. Free, without the usual catches
No account, no watermark, no trial, no paid tier and no cap on the number of videos. In our caption guide, the free plans of CapCut, VEED and Kapwing put watermarks on at least some exports, Submagic offers a trial rather than a free plan, HappyScribe's free captioning is a one-time 10-minute trial, and Clipchamp's free plan needs a Microsoft account.
3. Finishing tools for the caption itself
Word-level timing, pacing presets from one-to-three-word captions to two-line subtitles, five burn-in styles with optional animation and a highlight color, and four output formats. The editor keeps timing when you correct a line, as the edit we made showed in the exported file. For a single-purpose tool, that covers the caption job from transcript to finished file.
Who it is for
- People captioning footage they cannot upload: client interviews, internal training, school and university video, casework recordings. The video stays on the device, which the code confirms.
- Anyone who is not on a Mac and wants local transcription without installing software. It runs in a current browser on a computer, and the vendor says on an iPhone or an Android phone too.
- Short-form creators who want burned-in captions with styles and animation, and no account or watermark.
- YouTubers and editors who need an SRT or VTT file. The vendor notes that YouTube Studio, Premiere Pro, DaVinci Resolve, Final Cut Pro and CapCut all take SRT.
Not a good fit for
- Chinese, Japanese, Thai, Lao, Burmese, Khmer or Tibetan speech. These are not supported.
- Translation, speaker labels or batch captioning. None is offered.
- Long recordings on a phone or tablet. The 1-hour cap and the smaller model both apply.
- Organizations whose procurement needs terms of service and a named supplier. The site has neither.
- Difficult audio that must come out near-perfect. Larger models, such as those in MacWhisper Pro, or HappyScribe's human transcription are the safer route.
Key features
Transcription
- Whisper in the browser: small on computers, base on phones, tablets and low-memory devices, both 8-bit.
- 92 languages, with automatic detection or a manual choice.
- Voice activity detection that skips silent stretches, so nothing is transcribed from the gaps.
- A time for every word, so captions appear as the words are spoken.
Editing
- Click any line to correct it; the edit keeps its timing.
- Flagged words in English, according to the vendor: words that are often misheard are highlighted for checking.
- Pacing presets from one-to-three-word captions to two-line subtitles.
Styling and export
- Five styles, Classic, Minimal, Bold, Highlight and Pill, with Bounce, Pop, Fade or Reveal animation, size, position and any highlight color.
- Four outputs: MP4 with burned-in captions, SRT, VTT and TXT.
Storage and privacy
- Resume after a reload: the video and the lines so far are kept in the browser for up to a day.
- Recent work: transcripts of the last five videos are kept in the browser for 14 days; Clear saved work removes them.
- Offline use once the page and model have loaded, according to the vendor.
Pros & Cons
Pros
- The video and transcript never leave your device, which the site's code confirms
- Free with no account, no watermark, no paid tier and no cap on the number of videos
- 92 languages with automatic detection
- Four outputs: burned-in MP4, SRT, VTT and TXT
- Word-level timing, silence skipping, pacing presets and five caption styles
- Click-to-correct editor; the fix we typed kept its timing
- Accurate on clear speech in our test: 2.2% to 2.9% word error rate on studio and interview audio
- Runs on computers and, per the vendor, on phones, and resumes after a reload
- Unusually specific privacy documentation, including the list of hosts a page contacts
Cons
- No Chinese, Japanese, Thai, Lao, Burmese, Khmer or Tibetan
- No speaker labels, no translation and no batch captioning
- Smaller Whisper models: small on computers, base on phones and tablets
- Misspelled the mission acronym a third of the time, and filled one pause with an invented sentence, in our 37-minute test
- CPU only by default, so long videos are slow
- First use downloads about 340 MB on a computer, 155 to 170 MB on a phone
- Burned-in MP4 depends on the browser's H.264 encoder
- Error reports are sent automatically, without an opt-in
- Very new, with no terms of service, no named operator and no independent reviews
What to watch out for
Error reports go out without asking
If a tool fails or stops partway, technical details about the file and your browser go to PostHog's EU cloud automatically. They carry no media and no lasting identifier, but there is no opt-in, and an opt-in would fit a privacy-first product better. According to the vendor, a network filter can block the PostHog address without breaking the tools.
Shared computers keep local copies
While captions are being made, the video sits in the browser's storage for up to a day, and the last five transcripts stay for 14 days. On a classroom or shared computer, press Clear saved work when you finish, or clear the site's data in the browser.
The first run is a large download
About 340 MB on a computer and 155 to 170 MB on a phone, once per browser. On a phone, use WiFi rather than mobile data.
Tablets and low-memory laptops count as phones
The page decides by input type and reported memory, not by screen size. A touch-first tablet, or a computer that reports 4 GB of memory or less, gets Whisper base and the 1-hour cap.
Burn-in depends on your browser
The MP4 is encoded by the browser. Where the browser cannot encode H.264, the export stops with a message, and the SRT is the fallback.
No terms of service and no named operator
No terms page exists, and the site names its founder but not a company. For personal use that changes little, because nothing is uploaded and no account exists. For a school, council or business that needs a supplier on paper, it is a gap to raise with the vendor before adopting the tool.
Is Oh So Video worth it?
It costs nothing, so the real question is whether it fits the job.
If your footage cannot or should not leave your device, and you are not on a Mac, it is the practical choice: of the tools in our caption guide, only MacWhisper also keeps the audio local, and it needs a Mac. The privacy claim checks out in the code, there is no account to create, and you get a finished MP4 or a subtitle file without a watermark. The same goes for anyone who wants burned-in captions for short-form video without signing up to an editor.
If you need translation, speaker labels, Chinese, Japanese or Thai, batch processing or a team workspace, it is the wrong tool, and the price does not change that. The cloud tools in our guide cover those jobs, at the cost of uploading your audio.
For everyone in between, accuracy decides it. On our 37-minute test the captions were good enough to correct rather than rewrite, weaker on speech recorded in a room than on studio audio, and shaky on unusual names and acronyms. Trying costs nothing: caption a two-minute clip of your own speech, read the result, and you will know whether the corrections are a quick pass or a rewrite.
One caution on longevity. This is a free product from a vendor that launched in September 2026, with no terms of service and no stated business model. That does not affect a video you caption today, but it is a reason not to make it the only tool in a workflow you depend on.
How we arrived at 4.1
The score is the weighted result of ToolChase's published 8-parameter framework, using the default weights on our methodology page: core product quality 25, ease of use 15, value for money 15, feature depth 15, reliability 10, integrations 8, market trust 7 and support 5. Each category is marked out of 5, the weighted total runs from 0 to 100, and the methodology table converts it to the 5-point score, so the number can be reproduced from the reasoning column.
| Category | Weight | Score | Reasoning |
|---|---|---|---|
| Core Product Quality | 25 | 4.0 | The pipeline is complete and confirmed in the site's code: a voice detector skips silence, Whisper writes a time for every word, the editor keeps timing on corrections, and burn-in offers five styles with pacing presets. On a 37-minute NASA podcast the word error rate was 4.6% against NASA's transcript (5.6% counting two archival clips it leaves out): 2.2% and 2.9% on studio and interview speech and 6.8% on recordings made inside a habitat, close to what our reviews of the cloud tools print for clear speech. Held at 4 because the mission acronym came out wrong a third of the time, one invented sentence appeared, phones get the smaller base model, and the test material had no strongly accented speech. |
| Ease of Use / UX | 15 | 4.0 | Nothing to install and no account: drop a file, and the language is detected for you. Click any caption line to fix it. Held at 4 by the first-run download of about 340 MB on a computer, CPU-only processing, which took 12.4 minutes from choosing the file to finished captions for a 37-minute episode in our test, and a phone tab that pauses when you switch apps |
| Value for Money | 15 | 5.0 | Everything is free: no account, no watermark, no paid tier, no trial and, per the vendor, no limit on the number of videos, with MP4, SRT, VTT and TXT all included. In our caption guide, Clipchamp's free plan also has no watermark but needs a Microsoft account and transcribes on Microsoft's servers, MacWhisper's free tier is Mac-only, and the rest add watermarks or are trials |
| Feature Depth | 15 | 3.5 | A complete caption toolkit for a single-purpose tool: 92 languages with detection, word-level timing, pacing presets, five burn-in styles with four animations and four export formats. Held at 3.5 because there is no translation, no speaker labels, no batch captioning, no timeline editor, and no support for Chinese, Japanese, Thai, Lao, Burmese, Khmer or Tibetan |
| Reliability / Performance | 10 | 4.0 | No server sits in the processing path, so there are no queues, rate limits or service outages once the model is cached. The 37-minute test file finished in one pass with no stall, going back over stretches of speech its first pass had skipped, and on the phone setting the page caught and cut two loops. Against that, transcription runs on the CPU only by default: transcribing the test file took 11.4 minutes on an M1 Pro laptop, about 18 minutes per hour of audio against the 10 the site states. A computer that runs out of memory falls back to the smaller model, and burn-in needs a browser that can encode H.264, with the SRT as the fallback. |
| Integrations / Ecosystem | 8 | 3.5 | No API, no direct publishing and no cloud storage connections. The integration story is standard files: SRT that, per the vendor, YouTube Studio, Premiere Pro, DaVinci Resolve, Final Cut Pro and CapCut import, VTT for web players, plus TXT and a finished MP4, from MP4, MOV, WebM or MKV input. Those formats reach every common editor and platform, so it sits above the middle despite having no connectors |
| Market Trust / Reputation | 7 | 2.0 | The claims we could check in code held up, apart from the Blur faces logger, which is now blocked, and the founder previously founded and sold the stock footage site Videvo, which third parties confirm. Held at 2 because the domain was registered in September 2026, the site names no company behind it, and we found no independent reviews or benchmarks |
| Support / Docs / Governance | 5 | 2.5 | Documentation is specific: a long FAQ, a plain-language privacy policy dated 17 October 2026 that matches the code, and an education page listing every host a tool page contacts. Support is a published email address plus an in-page bug report form. Held at 2.5 because there are no terms of service and no named operator, error reports are sent without an opt-in, and there are no admin controls for schools or businesses |
| Total | 100 | 4.1 | Weighted on the published scale from 0 to 100 and converted with the methodology table: 76.4 out of 100, which converts to 4.1. |
Read the shape of the table rather than the total. Value sits at the top because there is nothing to pay and nothing held back behind a tier. Core quality, ease of use and reliability sit at 4: the captions came out close to the cloud tools on clear speech, there is no account, upload or install, and the long test file ran cleanly, held back by smaller models, a misspelled acronym and CPU-only speed. Features and integrations land at 3.5: a complete caption toolkit in standard file formats, without translation, speaker labels, seven languages or any connectors.
Market trust and support sit lowest, and both are about the vendor rather than the captions: a product launched in September 2026, with no independent reviews, no terms of service, no named operator, and error reports that go out without an opt-in.
For calibration against the caption tools already scored here: OpusClip 4.6, MacWhisper 4.5, CapCut, Submagic and VEED 4.4, Captions, HappyScribe and Kapwing 4.3, and Clipchamp 4.1.
At 4.1, Oh So Video sits level with Clipchamp and below the rest of the guide. The gap comes from market trust, governance and features rather than from the captions themselves, which came out close to what the guide's cloud tools promise on clear speech.
What would move the number: published terms of service and a named operator would lift market trust and support; an opt-in for error reports would lift support; a graphics-chip path for ordinary visitors would lift reliability and ease of use; translation or speaker labels would lift features. None of these would change the privacy design that is the reason to use it.
📋 Good to know
Open the captions generator in a current browser, drop in an MP4, MOV, WebM or MKV file, and pick a language or leave it on automatic. The first run downloads the speech model.
The video and transcript stay on your device. When a tool fails, an automatic error report with the file type, a rough size and browser details goes to PostHog's EU cloud. Visits are counted without cookies.
Free. No paid tier, no account and no watermark.
Low. Drop a file, correct lines by clicking them, choose a pace and a style, and export.
🔄 Alternatives by use case
Explore more
FAQ
Is Oh So Video free?
Yes, completely. There is no account, no trial, no paid tier and no watermark on anything you export, and the vendor states there is no limit on how many videos you caption. The only limits are per video: up to 3 hours on a computer and 1 hour on a phone. The vendor says the tools cost almost nothing to run because the work happens on your device, and names no other business model.
Does Oh So Video upload my video?
No. The captions tool reads the file, decodes the sound and runs the Whisper speech model inside your browser, and we found no code that sends the video, its audio, the transcript or the file name anywhere. Some technical data is sent: if a tool fails or stops partway, an automatic error report with the file type, a rough size band, the error and your browser details goes to PostHog's EU cloud, without an opt-in. Visits are counted by Plausible and Cloudflare without cookies.
How accurate is Oh So Video?
Accurate enough to correct rather than rewrite. We captioned a 37-minute NASA podcast and compared it with the transcript NASA published: the word error rate was 4.6%, with 2.2% for a host recorded in a studio, 2.9% for an interview and 6.8% for crew recorded inside a habitat, or 5.6% overall if two archival clips the transcript leaves out are counted. The mission acronym came out wrong a third of the time, and once the model invented a 14-word sentence in a pause, so read the captions through before publishing. That test ran on a computer, which gets Whisper small. On the phone setting, which uses the smaller Whisper base, the same episode came out at 7.4%.
Which languages does Oh So Video support?
92 languages, with automatic detection from the first half-minute of speech or a manual choice. Seven languages are excluded on purpose: Chinese, Japanese, Thai, Lao, Burmese, Khmer and Tibetan, which the vendor says is because they are written without spaces between words. Captions come out in the language spoken. The tool does not translate.
Is there a length or file size limit?
There is no file size limit. Length is capped at 3 hours per video on a computer and 1 hour on a phone, and the phone cap also applies to tablets and to any device that reports 4 GB of memory or less, because the page treats those as phones. The first use downloads about 340 MB of model files on a computer and about 155 to 170 MB on a phone, so use WiFi on a phone.
What formats can I export?
Four: an MP4 with the captions burned into the picture, SRT and VTT subtitle files with the same text and timings, and a plain TXT transcript. Burned-in captions come in five styles with optional animation, size, position and highlight color. The MP4 is encoded by your browser, so a browser that cannot encode H.264 video cannot make it; the page then points you to the SRT, or to making the MP4 on a computer or in another browser.
Does Oh So Video work on a phone?
Yes, in an up-to-date browser on an iPhone or an Android phone, according to the vendor. Phones get Whisper base, a smaller model than the Whisper small that computers get, and accuracy drops: on our 37-minute test the phone setting came out at a 7.4% word error rate, against 4.6% on a computer. Videos are capped at 1 hour. Keep the page open while it works, because switching apps pauses it.
Should I use Oh So Video or MacWhisper?
Both run Whisper on your own hardware, so neither uploads your audio. MacWhisper is a Mac app: its Pro license, €65 one-time, adds the larger Large v3, Turbo and Medium models, batch transcription and speaker diarization, and it exports TXT, SRT, VTT, CSV and DOCX. Oh So Video runs in a browser on any computer or phone with nothing to install, uses Whisper small and base, and adds styled burned-in captions. On a Mac with difficult audio or several speakers, choose MacWhisper. On any other device, or for burned-in social captions, choose Oh So Video.
Related AI caption tools
Caption generator guide →MacWhisper
local private transcription app for Mac that runs OpenAI…
Clipchamp
Microsoft free video editor with AI features.
CapCut
Free video editor by ByteDance with AI-powered tools
Veed
browser-based AI video editor with auto-subtitles…
HappyScribe
AI transcription, subtitles, and meeting notes in 150+ languages with a human-accuracy option
Kapwing
collaborative AI video editor with auto-subtitles, smart…
📝 Report incorrect info about Oh So Video
Is Oh So Video the right pick for your use case?
Answer four questions and get a ranked shortlist for the job you actually have, scored against verified feature grades, with the reasoning shown for every tool. No signup. Sponsorship never changes the order.
Build my AI video stack →