Beyond transcription: how AI turns speech into writing

Here's a sentence someone might say while thinking out loud:

"okay so for the quarterly update I want to, um, lead with the revenue number, actually no, lead with the customer story first and then the revenue, because, you know, it lands better"

And here's what they meant to write:

"For the quarterly update, I want to lead with the customer story, then the revenue number, because it lands better."

Every word in the second version came from the first. Yet one is a transcript and the other is writing. The difference between them is what Wisper does, and it's a harder problem than it looks.

Speech and writing are different languages

We tend to think of speech and writing as the same language in two formats. Linguists will tell you otherwise. Spoken language evolved for real-time, face-to-face communication, where tone, timing and shared context carry a lot of meaning. Writing evolved for readers who can't interrupt you to ask what you meant.

So speech naturally includes things writing doesn't:

• Fillers ("um", "uh", "like", "you know") that hold your place while you think.
• Self-repairs ("lead with the revenue, actually no, the customer story") where you change course mid-sentence.
• Repetitions ("we should, we should probably").
• Prosody, the rhythm and intonation that does the job punctuation does on the page.
• Loose structure, with ideas arriving in the order you think of them, not the order a reader needs.

A transcript captures the words but loses the prosody and keeps everything else. That's why raw transcripts are tiring to read.

Step one: hearing the words

Everything starts with automatic speech recognition (ASR), which converts audio into a sequence of words. Modern ASR systems are trained on vast amounts of speech and are remarkably accurate on clear audio. They still face real challenges:

• Names and jargon. Your colleague's surname, your product's codename or a niche technical term may not appear often in training data.
• Homophones. "Their", "there" and "they're" sound identical. So do "write" and "right".
• Noisy environments and accents the model has seen less often.

This is why Wisper lets you teach it your vocabulary. A short list of names and terms you use makes a real difference to accuracy where it matters most.

Step two: understanding what you meant

Once we have the words, the interesting work begins. This is where language models come in: they read the transcript as a whole and work out the intended message behind it.

Removing disfluencies, carefully

Deleting every "um" is easy. Knowing what else to remove is not. Consider:

• "We should, we should ship Tuesday" → the repetition goes.
• "I really, really need this by Friday" → the repetition is emphasis, so it stays.
• "Lead with revenue, actually no, the customer story" → the first option was abandoned and should disappear completely.
• "Like I said, the numbers are fine" → "like" isn't a filler here.

Getting this right requires understanding meaning, not matching patterns.

Punctuating by meaning

Basic dictation often inserts punctuation based on pauses. But people pause to think in the middle of sentences, and race through sentence boundaries when they're excited. Wisper punctuates by sentence structure and meaning, so commas, full stops and question marks land where a careful writer would put them.

Restoring structure

Spoken thoughts often arrive out of order or run together. Where it helps, Wisper breaks a long spoken passage into paragraphs, turns a spoken list into a written list, and smooths transitions, while keeping your ideas in the order you chose.

Step three: keeping it yours

This is the part we care about most. The fastest way to make cleanup worse is to make it too clever, rewriting your words into something generic, polished and unrecognisable.

So Wisper follows a few firm rules:

• Preserve meaning. Cleanup must never add claims, change facts or soften a point you meant to make.
• Preserve voice. If you write casually, the output should sound casual. If you're formal, it should stay formal.
• Preserve specifics. Names, numbers, dates and technical terms are kept exactly.
• Prefer minimal edits. When in doubt, change less.

The result should read like you on a good day, not like someone else entirely.

Why speed matters here too

All of this has to happen fast. If you finish speaking and then wait, the magic disappears. Wisper is designed so clean text arrives moments after you stop, quickly enough that dictating feels as immediate as typing, just faster.

Where AI cleanup still needs you

We're proud of what modern models can do, but we're honest about their limits. Cleanup can occasionally misjudge an ambiguous phrase or mishear an unusual word. That's why we recommend a quick read before you send anything important. It's still dramatically faster than typing from scratch, and it's the same habit you'd apply to anything you write.

From transcript to writing

The difference between dictation that people abandon and dictation that people rely on isn't the speech recognition. That problem is largely solved. It's everything that happens after: turning how we talk into how we write, without losing what makes it ours.

That's the problem Wisper is built to solve. If you want to try it, join the waitlist.