Why Two Transcript Tools Give Different Answers for the Same Video

You paste a link into one transcript tool and it says the video is unsupported. You paste the identical link into another and get a transcript in eight seconds. A third accepts it, works for two minutes, and fails.

The natural conclusion is that one tool is better. Sometimes that is true. Much more often the disagreement comes from four mechanisms that have nothing to do with quality, and knowing them tells you which tool to reach for next instead of trying all of them.

1. The platform is answering the two tools differently

This is the big one and it surprises people, because we are used to thinking of a URL as a fixed thing that either works or does not.

Several video platforms decide what to serve based on who is asking. The signals include the network the request comes from, the exact shape of the TLS handshake, how many requests that source has made recently, and whether the client looks like a browser a real person would use. Two tools running from different networks get genuinely different answers to the same question at the same moment.

We can show this happening rather than assert it. The same link, probed three times in a row from one machine on 15 September 2026: answered, answered, then HTTP 403 Forbidden. Nothing about the video changed. The wall closed because of how many times we had asked.

So a tool that "does not support" a platform may support it perfectly well from a different address, and a tool that worked for you this morning may be refused this afternoon. If your own repeated attempts change the outcome, you are looking at a rate limit, not a capability.

2. They disagree about what to download

To turn speech into text you need the audio and nothing else. Most platforms publish an audio-only version, so a tool asks for that and downloads a few megabytes.

Some platforms publish no audio-only version at all — every option is picture and sound welded together. A tool that asks for "the best available" then downloads the largest file on the site to reach audio that is bit-for-bit identical in the smallest one. We have measured that penalty at 123 times on one platform and 22 on another.

Both tools succeed on a five-minute clip. On a ninety-minute one, the tool asking for the smallest usable version finishes and the tool asking for the best times out. Same link, same platform, opposite outcomes, and the difference is one line of policy neither tool shows you. Why a transcript runs for two minutes and then fails is the long version.

3. They disagree about how long to wait

Every tool has a patience limit, and most scale it to the length of the video. That works until the platform refuses to say how long the video is — which, per mechanism 1, happens intermittently on exactly the platforms with walls in front of them.

What a tool does with an unknown length is an invisible policy decision with visible consequences. Treat it as "probably short" and a long video gets killed early with a misleading message about a slow download. Treat it as "could be anything" and it finishes. We had this backwards until 15 September 2026: a job whose length we could not read was given 120 seconds when, had the probe answered, it had earned 479.

4. One of them is being honest about a wall it cannot pass

Some videos cannot be transcribed by any tool, because access is the point. A video behind a login, a paywall, or a region lock is not a technical puzzle — somebody is deliberately not serving it to strangers.

Here tools differ in candour rather than capability. One says "unsupported". One says "please try again", which is worse than useless because retrying cannot work. One tells you which wall you hit and whether anyone can do something about it. Which restricted videos can actually be transcribed splits the six cases; one needs nothing but the full link, two need the owner to change a setting, and three never work.

So which tool should you use?

A practical order, and it costs nothing to follow:

1. If the video has captions on the platform, almost anything works, including free browser extensions. Do not pay for this case.

2. If it does not, the question is whether the tool handles that platform's audio well — which you cannot read off a feature list. The honest signal is whether the tool publishes how often each platform actually succeeds. Ours is here, including the platforms we are bad at.

3. If one tool fails, try again before switching. Given mechanism 1, the second attempt on the same tool is often a better bet than the first attempt on a different one.

4. If every tool fails the same way, believe them. That is mechanism 4, and no tool is going to out-engineer a paywall.

The uncomfortable part

Most of these differences are invisible to you. Nothing in a tool's interface tells you which format it asked for, how long it was prepared to wait, or whether the platform served it a refusal dressed up as a page.

That is the real argument for a tool publishing its failure modes rather than its feature list. A supported-platform list costs nothing to write and tells you nothing. A success rate, an error message that names the actual wall, and a refund when the job dies are all falsifiable — which is precisely why they are worth more.