What We Actually Measured: Transcript Success Rates for 21 Video Platforms

Every transcript tool advertises a list of supported platforms. Almost none publish how often those platforms actually work, which makes the lists close to meaningless — a site can be "supported" and still fail two times in five.

So here are ours, measured rather than claimed. As of 15 September 2026 we have run 16,286 transcript jobs, of which 14,778 reached a final state: 12,403 finished and 2,375 failed. That is an 83.9% all-time success rate across 3,870 hours of video and 6.1 million transcript segments, for 8,240 people.

The remaining 1,508 are quotes nobody confirmed — somebody pasted a link, saw the price, and closed the tab. They are not failures and they are not counted as attempts.

The last 30 days, by platform

This is the number that matters if you are choosing a tool today. Platforms with fewer than 15 attempts in the window are left out, because a rate built on eight tries is not a rate.

• Brightcove — 100% (18 attempts)
• SoundCloud — 100% (43)
• Uploaded files — 100% (49)
• Vidyard — 100% (26)
• Bilibili — 97.4% (38)
• Apple Podcasts — 96.9% (64)
• YouTube — 95.8% (263)
• Wistia — 93.8% (65)
• Vimeo — 93.6% (672)
• Instagram — 93.0% (200)
• X / Twitter — 92.6% (27)
• Facebook — 91.8% (1,117)
• Twitch VODs — 91.2% (68)
• Dailymotion — 88.8% (98)
• Kick — 88.2% (51)
• LinkedIn — 86.7% (98)
• Google Drive — 76.1% (205)
• Odysee — 75.0% (24)
• Rumble — 72.2% (133)
• TikTok — 61.6% (86)

Why the all-time numbers are worse, and why that is the honest comparison

Over the full history the same platforms read very differently: Kick 42.1%, Dailymotion 68.6%, Vimeo 80.0%, TikTok 63.6%, YouTube 87.4%.

The gap is not seasonality. It is repair work landing. Kick went from 42.1% to 88.2% because its livestream VOD links resolve through an identifier that every metadata endpoint used to reject, so multi-hour recordings were priced and budgeted as if they were three minutes long and timed out. Dailymotion, Wistia and Rumble were all losing long videos to the same class of bug: those platforms publish no audio-only version, so the download kept reaching for the largest file on offer to get audio that was byte-identical in the smallest one. Vimeo recovered when logged-out playback stopped being available and we stopped pretending otherwise.

We publish both numbers because quoting only the good one would be the sort of claim this post exists to argue against. If you are evaluating us, the 30-day column is what you will experience. The all-time column is the honest record of getting there.

The platforms we are still bad at, and why

TikTok, 61.6%. This is a wall, not a bug. TikTok's edge refuses requests that do not look like a real app, and the refusal is not stable — the same link can fail, then work thirty seconds later, then fail again. We have measured 0 out of 12 across three different extractor builds on a day when the platform was simply refusing everyone. Things that did not help, each tested and discarded: TLS fingerprint impersonation, routing through a residential proxy, and the mobile API, which needs a device install identifier we have no legitimate way to hold. There is more detail in why a TikTok transcript fails and then works.

Rumble, 72.2%. Also a rate wall, and a well-behaved one: probe the same eight links repeatedly and the success rate falls as you probe, with individual links flipping in both directions. Impersonating a Safari or Firefox handshake gets through where Chromium profiles do not, which is why the number is 72% rather than single digits.

Google Drive, 76.1%. Almost none of this is ours. A Drive file has a sharing setting, and the common failure is a link shared with specific people rather than with anyone holding the link. That is fixable in about ten seconds by the person who owns the file, and it is the one failure in this list where the message tells you exactly what to change.

Odysee, 75.0%. Small sample, 24 attempts. Say the n out loud or the number is decoration.

What to take from this

Three things, none of which are specific to us.

1. A supported-platform list is not a success rate. If a tool lists a platform, ask what fraction of those jobs finish. Nobody who is winning that comparison keeps it secret.

2. Failure rates cluster by mechanism, not by luck. A platform sitting near 50% usually is not flaky — it is deterministically failing one category of video, most often long ones. That pattern is recognisable and fixable. A platform sitting near 90% with scattered failures is usually meeting a rate limit.

3. Ask what happens to your money when it fails. Ours refunds automatically and the ledger is checked for it; see what transcription actually costs.

If you want to know whether one specific link will work before spending anything, six checks that take ten seconds is faster than reading a table. If you have already hit an error, every transcript error and what it means has the exact wording.

Method, so you can discount it properly

Numbers come from our own production database on 15 September 2026. An attempt is a job that reached completed or failed; unconfirmed quotes are excluded. Success means a transcript was produced and stored, not that it was good — accuracy is a separate question covered in how accurate AI transcription actually is.

The obvious bias: these are jobs real people chose to run, from one set of servers in one country. A platform that geo-restricts will look worse to us than it would to you somewhere else, and our per-platform samples range from 18 attempts to 1,117. Weight them accordingly.