The short answer
Speech recognition does not identify words purely from sound. It combines what it heard with what is statistically likely to come next, and that second part is what rescues it from noisy audio. A proper noun has almost no statistical support — your co-founder's surname appears nowhere in the model's sense of what usually follows "and this is". So the model falls back on the nearest common word that fits the sound, confidently.
The practical consequence is useful: errors are predictable. They concentrate on names, brands, places, numbers and field-specific terms. If you review those categories first instead of reading every caption evenly, you find most of your errors in a fraction of the time.
What 95% accuracy actually feels like
Clean English audio — good mic, one speaker, quiet room — gets you around 95% of words correct. That number reads like a solved problem and is not.
Count it out. People speak at roughly 150 words a minute, so a one-minute clip is about 150 words and 95% leaves seven or eight wrong. A ten-minute video has seventy-five errors in it.
Then the distribution makes it worse. If those errors were spread evenly across common words, most would be harmless — a dropped "the" nobody notices. They are not spread evenly. They land on exactly the words carrying your meaning and your credibility: your own name, your product, the number in your central claim.
And 95% is the ceiling, measured on clean audio. Clean Hindi is around 85-90%. Add background music, a second speaker, a phone mic at arm's length or a reverberant room and every figure drops.
Why the model is guessing at all
It helps to know what is happening, because it tells you where to look.
A speech model is doing two things at once: mapping sound to plausible words, and weighing those candidates against what makes sense in context. The second part is why recognition works at all on real audio. Human speech is full of ambiguity, slurring and noise, and the acoustic signal alone is frequently not enough to distinguish two similar-sounding words. Context resolves it.
This works beautifully for common language and collapses for proper nouns. The model has seen "recognise speech" in training and can prefer it over "wreck a nice beach" from identical sound. It has never seen your surname, so there is nothing to prefer. It picks whatever common word fits the acoustics, and it does so with no signal that it was unsure.
Which is why the errors feel so pointed. The model is not failing randomly — it is failing specifically wherever the language is unusual, and the unusual words in your video are the ones that are specifically yours.
The five categories, in order
Review in this order and you will find most errors quickly.
Proper nouns first. People, companies, places, product names. Highest error rate and highest cost, since your audience spots a misspelled name instantly and it reads as carelessness about your own subject.
Then cross-language words. English terms inside a Hindi sentence, or any code-switching. The model has to decide which language it is hearing mid-phrase and which script to write in, and either decision can go wrong independently.
Then numbers. Prices, dates, percentages, quantities. Moderate error rate, severe consequences — a misheard digit in a factual claim is a different claim. They are also easy to skim past, because a wrong number still reads as a plausible sentence.
Then domain terms. Your field's vocabulary is rare in general training data. A mis-transcribed technical term in an educational clip is worse than an obvious error, because it reads as authoritative and a learner cannot tell.
Last, homophones and near-homophones. These need context to resolve and the model sometimes resolves them wrongly. Lower priority, because they are usually obvious when you read the line.
Review with the video playing
Method matters as much as order. The default — a transcript in a text box next to a video player — imposes a tax on every single check. For each suspicious line you locate the moment in the video, play it, decide, and come back. The keystrokes are not the cost; the context switch is, repeated dozens of times per clip.
Reviewing with the captions over the playing video removes that entirely. You watch your clip as a viewer would, and when something is wrong it is wrong in front of you at the moment it is spoken. You fix it there.
It also catches a class of problem a transcript physically cannot show you: a caption that is perfectly accurate but on screen too briefly to read, or positioned where it covers something that matters. Those are caption bugs rather than transcription bugs, and they are invisible in a text box.
Reducing the errors before they happen
The highest-leverage change is not software. It is your microphone.
Get the mic close to the speaker. A clip-on lavalier or a phone held at a sensible distance beats a laptop microphone across a room by a wide margin, because proximity raises the signal relative to the room.
Record in a space with soft surfaces. Reverberation smears the acoustic boundaries between words, and hard parallel walls are the worst case. Carpet, curtains and furniture all help.
Do not put music under speech during recording. Add it in the edit, after captioning, so the captioning pass hears clean voice.
One speaker at a time. Overlapping speech is among the hardest problems in the field and crosstalk degrades both voices.
Say unusual names clearly and slightly slower. Those are the words with no statistical safety net, so the acoustic signal is all the model has to go on.
None of this is glamorous, and all of it will cut your correction time more than any feature in any captioning tool — because it moves the audio toward the top of the model's range where there is simply less to fix.