How speech to text actually works (the words, not the captions)

The search box sold you captions

Type audio transcription or transcribe audio to text and the pages that come back treat every recording as one job: upload a file, get a document, export SRT, maybe color the speakers. That is a clean demo. It is a bad map if you needed the words from an interview, a chapter take, or a voice note.

Most of the work we see is not a caption file. It is a transcript: a string of words you can search, quote, edit, and later re-voice. Captions need times. Meeting notes need who said what. Push-to-talk dictation needs the text at your cursor in another app. Those are different jobs. Speech to Text does the first one, and we will not advertise the others until they are actually here.

The short definition lives in what is speech to text. This post is the craft version: the pipeline you can hear fail, the three-job split, and the limits we print on purpose.

Three jobs, not one upload

People type the same phrase for three products. Mixing them up is how you end up with a wall of text and no way to cut it to picture, or a dictation app that cannot open last Tuesday's mp3.

What the model is doing

You give the system audio of someone speaking. It returns written words. That is the whole product promise. Under the hood the recording is turned into a picture of loudness across frequencies over time (a spectrogram). A neural model reads that picture and writes text.

Older stacks split the work into an acoustic model, a pronunciation dictionary, and a language model. Modern Whisper-class systems fold most of that into one network. You still experience the same sequence: audio in, text out, mistakes that look like confident English.

Because the decoder is also a language model, it can insert punctuation and pick their versus there from context. That is why a good transcript reads like a paragraph instead of a stenographer dump. It is also why silence is dangerous. When the audio is empty or muddy, the language side will sometimes invent a plausible sentence: a thanks for watching, a repeated line, a clause nobody said. That is a hallucination, not a quirky accent.

File transcription here

The live path is Speech to Text. Upload a recording or capture one in the browser. A Whisper-class model returns an editable transcript, usually in seconds. The studio header labels the model class Whisper v3 Turbo so you can see what produced the text in front of you.

Accepted uploads: mp3, wav, m4a, webm, ogg, and flac, up to 25 MB per file. The mp4, mpeg, mpga, and oga extensions are read too; for mp4 we take the audio track. Browser recording is capped at five minutes per take. Compressed formats are fine. What hurts is room echo, two people talking at once, and music under the voice.

The transcript comes back as plain text. Copy it, download a .txt named after your audio, or save the source file with its transcript to your library (library save needs sign-in). There are no speaker labels and no timestamps, because the endpoint does not return them, and we will not fake structure it did not give us. The longer how-to is how to transcribe audio to text.

Honest metering: we count the transcript's characters against the same monthly allowance the voice tools spend, not the audio's wall-clock. About 1,000 characters is about a minute, so the two views stay close in practice. A failed or empty job costs nothing.

Dictation is a different door

If the job is "talk, and have the sentence land where I am already typing," that is not an upload. Scribe is a small Windows app: hold the hotkey (Ctrl+Space by default), speak, release, and the transcript is pasted at your cursor in whatever app has focus. A document, an email, a chat box, a code comment.

It uses the same disclosed transcription family as the studio tool. Audio is held in memory on your machine, never written to disk. Holds are capped at sixty seconds per take in this beta. The installer is not code-signed yet, so Windows SmartScreen asks once; that warning is expected while signing is in progress. There is no Mac or Linux build today. Do not download the x64 installer onto a Windows-on-ARM machine and expect it to run.

Scribe is dictation. Speech to Text is file transcription. If you mix those doors you will fight the product instead of finishing the page.

Where it still fails

Accuracy is mostly decided before the model runs, by the recording. Timbre of the original voice can be excellent while the transcript is still unusable. These are the misses we see when people treat the search phrase as a spec.

• Distance and echo: a speaker far from the microphone in a live room.
• Overlap: two people talking at once is genuinely hard.
• Music or a bed under the voice.
• Names, brands, and jargon the model has never seen spelled.
• Silence or long gaps: invented sentences that sound finished.
• A wall of text treated as a caption file, then blamed when it will not cut to picture.
• A two-person interview treated as minutes, then blamed for missing speakers.

After the words come back

Read the names before you trust the file. Proper nouns are where a fluent transcript quietly dies. Fix them in the text. Do not re-run the audio ten times hoping the spelling appears.

If the point of the transcript was a new performance, take the edited words into Text to Speech and generate. That is a second job with a second set of engines, not a hidden switch on the transcribe button.

If the point was another language, transcription is step one of Dubbing & Translation. We already wrote the localize craft in how AI dubbing actually works. Do not skip the inspectable script. A bad name in the transcript becomes a bad name in the new language.

If the point was captions, stop. Words without times are not a subtitle file. Inventing stamps in a spreadsheet is how you ship late captions that drift. We would rather print a later than export fake SRT. The short split is dubbing vs subtitling.

Honest limits

No speaker diarization. No word or segment timestamps. No subtitle export. When the endpoint gains timing, the docs will change with it.

Uploads stop at 25 MB. Browser takes stop at five minutes. Scribe holds stop at sixty seconds, Windows only, unsigned installer for now.

We do not publish an accuracy percentage we have not measured. When we do, the number will carry its date and method, the same standard as every figure on the open benchmark.

Transcription spends the same flat monthly allowance as voice (published free-tier budget: 10,000 characters, about 10 minutes by house math). Paid plans raise the cap. They do not put you on a per-character anxiety meter.

This tool does not make voices, clones, or music beds. Speech is Text to Speech. Instrumental underscoring is Sound & Music.

Transcribe one this week

The shortest proof is a recording you already have, not a roundup of vendors.

• Pick a single-speaker take if you can: a voice note, a chapter, an interview where one person mostly talks.
• Open Speech to Text, upload it (or record a short take in the browser), and wait for the string.
• Mark every name, number, and brand. Fix those in the text.
• If you needed captions, notice the missing times and stop. Do not invent them.
• If you needed the sentence in another app while you talk, that is Scribe, not this upload box.
• If you needed a new read, paste the cleaned words into Text to Speech.

Open the studio, not the magic caption

If you came here from an audio-to-text or Whisper-transcription search, the honest product is a plain-text transcript of a file you already have. Open Speech to Text, or start from the Speech to Text guide if you want the short version first.

When you want the full studio, create a free account. Written to be heard still starts with getting the words right, as long as we do not lie about what those words are missing.