Free speech to text converter that runs in your browser

What speech to text can and cannot do for you

Transcription software listens to sound and guesses at words. That word, guesses, is doing real work in the sentence. A model does not understand what you are saying. It has learned which sequences of sounds usually correspond to which sequences of letters, and it produces the most likely answer it can find.

On clear speech that produces something remarkably close to what was said. On a recording made across a meeting room table with three people talking over each other, it produces something you will need to sit with. Knowing where the line falls saves a great deal of frustration.

What it reliably gives you is a searchable first draft in a fraction of the time. Transcribing an hour of audio by hand takes most people four to six hours of stopping, rewinding and typing. Getting a rough transcript in minutes and then spending twenty minutes fixing it is a different kind of afternoon.

The two routes on this page, and why they are not the same

Most transcription sites offer one method and describe it vaguely. There are two here and the difference between them is worth understanding before you choose.

Transcribing a recording

A speech model is downloaded to your browser once and then runs inside the tab. Your audio file is opened, decoded and processed on your own machine. Nothing about the recording is sent anywhere. The trade is a one time download of forty to eighty megabytes and a little patience on the first run.

Dictating with a microphone

This uses the speech recognition built into your browser. It is instant and needs no download, which makes it excellent for quick notes and reminders. It is also, in Chrome and Edge, not private. Those browsers work by streaming the audio to Google servers to be recognised and sending the words back.

We say that on the page itself rather than in small print, because plenty of sites offering a microphone button describe themselves as completely private while quietly using exactly the same mechanism. If what you are saying should not leave your desk, record it and use the file route instead.

Where people use this

• Interviews. Journalists, researchers and recruiters who need quotes they can search rather than a two hour file they have to scrub through.
• Lectures and lessons. A recorded class becomes notes you can actually revise from, with the words on the page instead of somewhere in the audio.
• Meetings. Turning a recording into text makes actions and decisions findable weeks later when nobody remembers who agreed to what.
• Voice notes. Ideas captured while walking or driving, converted into something you can paste into a document.
• Accessibility. Transcripts make audio content usable for people who are deaf or hard of hearing, and they are the starting point for captions.
• Content work. Podcasters and video makers who want show notes, subtitles or a written version of what they recorded.

What decides how accurate the result is

Almost everything comes down to the recording itself. The model can only work with the sound it is given.

The things that always need checking

Even on a clean recording, four categories reliably come out wrong, because the model has to guess at words it has rarely or never met. Personal names, company names, acronyms and technical jargon. If your transcript matters, read those parts against the audio before you quote anything. A misheard surname in a published article is a bigger problem than a hundred small typos.

Recording so the transcript comes out clean

Five minutes of thought before you press record saves an hour of correcting afterwards.

• Get the microphone close. Distance is the single biggest factor. A phone lying on the table between two people is far worse than a phone held near whoever is speaking.
• Choose a room with soft surfaces. Curtains, carpet and furniture absorb the reflections that turn speech into mush.
• Turn off anything humming. Fans, air conditioning and fridges produce steady noise that survives compression perfectly, which is exactly what you do not want.
• Ask people not to talk over each other. Awkward to request, but crosstalk is where transcripts fall apart entirely.
• Have everyone say their name at the start. No transcriber here separates speakers, so a spoken introduction gives you an anchor when you label the text afterwards.
• Test for ten seconds and listen back. Finding out afterwards that the microphone was covered is a familiar and painful experience.

Choosing between quick and careful

Two models are offered and the choice is a straight trade between speed and accuracy.

Quick is around forty two megabytes and gets through audio faster than real time on an ordinary laptop, so an hour of recording finishes in well under an hour. It is the right choice for clear speech from a single speaker, which is most voice notes, most lectures and most one to one interviews.

Careful is roughly twice the size and takes noticeably longer, but it copes better with accents, background noise and overlapping voices. Reach for it when the quick model has produced something you are spending too long fixing. The download happens once for each model and is then cached by your browser, so trying both costs you the download only the first time.

Both models here are English only, and they are deliberately the smaller ones so that the page stays usable on ordinary hardware and average connections. Feeding them another language does not produce an error. It produces confident nonsense, which is worse, so match the model to the audio.

Working with a long recording

Long audio is cut into short passes before it is transcribed, because these models work in windows of about half a minute. The cut points are not placed blindly every thirty seconds. The tool looks for the quietest moment near each boundary and cuts there, so a pass rarely lands in the middle of a word.

Text appears pass by pass rather than all at the end, which means you can start reading and correcting while the rest is still running. Turning on timestamps marks each pass with the time it starts, which is what makes a two hour interview navigable afterwards. Find the quote in the text, note the time, jump straight to it in the audio.

Very long files depend on how much memory your browser will hand over. If a three hour recording will not finish, split it in half and transcribe each part. If you are starting from video, pulling the audio out first with our video to MP3 tool gives you a much smaller file to work with.

Turning a rough transcript into something usable

A raw transcript is not a document. It is raw material. A few passes turn it into something you would send to someone.

• Fix the names first, since they are both the most likely errors and the most embarrassing ones.
• Add paragraph breaks where the subject changes. Machines produce walls of text and humans cannot read them.
• Label the speakers if there is more than one, using the introductions you recorded at the start.
• Decide whether to keep the ums and false starts. Verbatim matters for research and interviews. For notes and articles, cut them.
• Read it once against the audio if anything in it will be published or quoted. Twenty minutes now beats a correction later.

Why running it in your browser matters here

Think about what actually gets transcribed. Interviews with sources who spoke on condition of anonymity. Therapy sessions and medical consultations. Legal discussions. Internal meetings about restructuring. Voice notes about things you have told nobody. This is some of the most sensitive material a person handles, and the usual model is to upload all of it to a company you know nothing about.

Those services store your audio, often on servers in another country, under retention terms almost nobody reads. Several have used customer recordings to train their models. That may be perfectly acceptable for your holiday video and completely unacceptable for a confidential interview.

The file route on this page never sends your