What is speech to text?

Speech to text, also called automatic speech recognition or ASR, is software that turns spoken audio into written words. It listens to the sound of speech, works out which words were said, and produces a text transcript. Modern versions run on AI models trained on thousands of hours of recorded voices, which is why they handle accents and background noise far better than the dictation tools of a decade ago.

What is the difference between live dictation and transcribing a recording?

Live dictation turns speech into text in real time as you speak, the way voice typing in a phone or word processor does. Transcribing a recording works on a saved audio or video file after the fact. Recordings usually come out more accurate because the engine can see the whole sentence before it commits to the words, and it can label who spoke when. Hushscript handles recordings, not live dictation.

How accurate is automatic speech to text?

For clear speech in a quiet room, current AI models reach roughly 90 to 97 percent word accuracy on well-supported languages. Accuracy drops with heavy background noise, two people talking over each other, strong accents the model has not heard much of, or specialist vocabulary like drug names or legal Latin. Cleaning up the recording does more for accuracy than any single setting.

Does speech to text work with accents?

Yes, and far better than older systems. Models trained on large, varied datasets handle US, UK, Australian, and Indian English comfortably, and Hushscript offers four distinct English accents. A strong regional dialect or a less-represented accent will need a little more correction afterwards, but the transcript is still a usable first draft rather than something you retype from scratch.

How do I convert recorded speech to text on Hushscript?

Drop your audio or video file on the audio-to-text page and a 30-second speaker-labeled preview appears with no account needed. Sign up with your email to transcribe the rest, then upload the full file. The finished transcript comes back with speaker labels and timestamps, and you can rename speakers and export to TXT, SRT, DOCX, or JSON.

Is speech to text free?

Hushscript is not free, because the AI behind it costs real money to run, so it is pay-as-you-go with no subscription. You do get 30 free minutes to try it. They unlock instantly when you validate a card with a $1 hold that is released right away and never charged, or with your first purchase if you pay another way. The 30-second preview and the in-browser tools are genuinely free with no account.

Which languages can Hushscript transcribe?

It detects the language automatically and transcribes around 99 languages, with top-tier accuracy in 18 of them, including Global, British, Australian, and US English, Spanish, French, German, Japanese, Mandarin, Hindi, and Arabic. The Languages page lists every one. A recording that stays in a single language transcribes best.

Speech to Text: How It Works and How to Do It

June 24, 2026

Speech to text is software that turns spoken audio into written words. The term covers two different jobs that feel similar but work differently: live dictation, where you speak and text appears as you go, and transcription, where you hand over a recording and get a finished transcript back. The core idea is the same in both. The workflow, the accuracy, and what you can do with the result are not.

This guide explains what speech to text actually is, in plain English, how the engine gets from sound to words, and how the two modes differ. Then it walks through converting a recording to text on Hushscript, with a worked example, the problems that trip people up, and a short FAQ at the end.

What speech to text actually is

At its simplest, automatic speech recognition (ASR) takes a sound wave and produces a sequence of words. A microphone captures speech as a continuous waveform, a wobbling line that represents air pressure changing over time. That waveform carries everything: the words, the speaker’s voice, the room, the hum of an air conditioner. The engine’s task is to pull the words out of that mix and write them down.

To do it, the engine chops the audio into tiny slices, usually 10 to 30 milliseconds each, and measures the sound in every slice. Those measurements feed a trained model that maps patterns of sound onto the speech units of a language, then onto words, then onto whole phrases. You do not see any of this. You see a recording go in and a transcript come out. But knowing roughly what happens inside explains why some recordings transcribe cleanly and others come back rough.

There are two layers of knowledge working together. One is acoustic: what the sounds of a language are and how they tend to be pronounced across different voices and accents. The other is linguistic: which word sequences are plausible. The linguistic layer is why a good engine writes “their car broke down” rather than “there car broke down,” even though the two sound identical. It is using context to choose the spelling a human would.

Older speech recognition relied on hand-written rules and small training sets, and it was fragile. Modern engines are large AI models trained on thousands of hours of recorded speech from many speakers, accents, and recording conditions. That breadth is the whole reason today’s transcripts are usable. On clear speech in a decent recording, current models land somewhere around 90 to 97 percent word accuracy. The remaining few percent is where the troubleshooting later in this guide earns its keep.

Recorded speech vs live dictation

The split that matters most in practice is whether the audio is being processed live or after the fact.

Live dictation runs in real time. You speak into a phone, a voice assistant, or the voice-typing feature in a word processor, and words appear almost as fast as you say them. To keep up, the engine has to commit to each word with very little of the sentence to lean on. That is a hard constraint, and it is why dictation sometimes guesses a word, then quietly rewrites it once the next few words arrive. Live dictation is the right tool when the point is to capture your own speech as you go: jotting a note hands-free, drafting an email by voice, controlling a device.

Transcription works on a complete recording. You give the engine a finished audio or video file, it processes the whole thing, and you get a full transcript in return. Because nothing has to happen in real time, the engine can read ahead and see the entire sentence before deciding on any word, which lifts accuracy. It can also do work that live dictation cannot. It can separate the speakers, attributing each line to whoever said it, and it can attach timestamps so you can jump from a line of text straight to that moment in the audio.

If you are dealing with a meeting, an interview, a lecture, a podcast episode, or a phone call you recorded, you are doing transcription from a saved file, not live dictation. The two get confused because both are marketed as “speech to text,” but the recording you already have is the wrong shape for a dictation tool and the right shape for a transcriber. Hushscript is built for recordings. There is no live-dictation mode, and that focus is deliberate: working from the full file is what lets it read ahead for accuracy and label speakers in the same pass.

How to convert recorded speech to text

Here is the actual flow on Hushscript, in the order you meet it. You need the recording as a file on your device, an audio or video file in any common format, and an email address to sign up. That is the whole list. There is nothing to install.

Drop the file and watch the preview. Open the audio-to-text page and drop your file onto it. Hushscript transcribes the first 30 seconds straight away and shows you the result with speakers already separated. No account is needed for this step. The preview is there so you can judge the accuracy on your own recording before you commit anything.
Sign up to transcribe the rest. If the preview looks right, sign up with your email. Transcribing the full file is gated, so this step is where an account comes in. New accounts get 30 free minutes to try the service properly.
Upload the full file. Once you are in, upload the whole recording. If it is a video, the audio is extracted in your browser first and only the audio is sent, so the video itself stays on your device.
Read, relabel, and export. The transcript comes back with each turn attributed to a speaker (Speaker A, Speaker B, and so on) and timestamped. Click any speaker label to rename it, and the new name updates everywhere that person spoke. When it reads the way you want, export to TXT, SRT for subtitles, DOCX for editing, or JSON for structured data.

Thirty minutes is enough to transcribe a short interview, a 10-minute lecture clip, or the first stretch of a longer recording if you want to check accuracy before going further. The free minutes unlock instantly when you validate a card with a $1 hold that is authorized then released right away and never charged, or with your first purchase if you use another payment method. A card is not required; it is just the fastest route to the bonus. After the free minutes, you pay only for the minutes you transcribe, with no subscription. The pricing page has the current rates.

A worked example

Say you recorded a 38-minute customer interview on your phone. It saved as an M4A, two voices, the occasional clatter of a coffee cup, recorded in a normal meeting room.

You drop the M4A on the audio-to-text page. Within a few seconds the 30-second preview appears, and it already looks like a conversation:

Speaker A   00:00   Thanks for making the time. Could you start by telling me
                    what first made you look for a tool like this?
Speaker B   00:11   Sure. We were drowning in support tickets and the old
                    system just couldn't keep up, so I started looking around.

That is the unedited preview. The two voices are split, each line is timestamped, and the wording is clean. You sign up, upload the full 38-minute file, and a couple of minutes later the complete transcript is ready in the same shape. Speaker A is you and Speaker B is your interviewee, so you click the “Speaker A” label, type your name, and click “Speaker B” to type theirs. Both rename throughout the whole document in one pass. You export to DOCX, open it in your word processor, fix the two or three places where a product name came out spelled phonetically, and you have a finished interview transcript. The 38 minutes came out of your balance, which is why the free 30 minutes are handy for a first real job rather than just a test clip.

Common problems, and how to fix them

Most disappointing transcripts trace back to the recording, not the engine. These are the issues that come up most, and what to do about each.

The recording is quiet or muddy. If voices are faint or the room sounds boomy, the engine has less to work with and accuracy falls. A microphone close to the speaker, a lapel mic or a headset within 10 to 20 centimetres of the mouth, makes a bigger difference than any setting. A large, bare room adds echo that smears the boundaries between sounds; a smaller or soft-furnished space records better. You cannot fix this after the fact, so it is worth getting right before you hit record next time.

Two people talk at once. When voices overlap, both the wording and the who-said-what get harder, because two sets of sounds are competing in the same slice of audio. There is no clean software fix for genuine overlap. In a recording you control, leaving a beat between turns pays off later. In one you cannot redo, expect a few crossed-over lines around the interruptions and tidy them by hand.

A specialist term comes out wrong. General-purpose models know everyday language well but have not necessarily seen your industry’s jargon, an unusual surname, or a product name. Those tend to be transcribed by how they sound rather than how they are spelled. A find-and-replace in the exported DOCX clears a recurring term in seconds, which is one reason DOCX is the easiest export to correct in.

A file will not load. Most standard audio and video files just work. If one refuses, it is usually an unusual or very old container. Converting it to a plain MP3 or WAV with the free in-browser tools, which run on your device and upload nothing, almost always solves it. If a format still will not go through, email support@hushscript.com and the team will add it.

One speaker gets split into two. Occasionally the same person is labeled as two speakers, normally because their voice changed partway through: they moved closer to the mic, switched from a headset to speakerphone, or the line quality shifted. The engine groups by how a voice sounds, so a big change can read as a new person. Merging the two labels in the editor fixes it in one step. The mechanics of this live in the guide on how speaker diarization works.

Accuracy and accents

Accent coverage depends on the model and the language. For English, models trained on broad datasets handle US, UK, Australian, and Indian English well, and Hushscript offers four distinct English accents. A heavy regional dialect or a less-represented accent variety will need a touch more correction, but you are still editing a draft rather than typing from nothing.

A few habits lift accuracy on any recording, whatever the accent. Record with a close microphone. Let each person finish before the next begins. Avoid big echoey rooms. A moderate speaking pace transcribes better than a sprint past 200 words a minute. And a reasonable bitrate matters: a 128 kbps MP3 is fine, while heavily compressed voice-message audio in narrowband loses detail the engine needs. For a fuller walk through the formats and the quirks of each, see how to transcribe any audio file.

Hushscript detects the language automatically and transcribes around 99 languages, with top-tier accuracy in 18 of them. A recording that stays in one language transcribes most cleanly; switching languages mid-sentence is hard for any engine.

Speaker labels, in the same pass

Separating the speakers, formally called speaker diarization, runs alongside the speech recognition rather than as a second job you pay extra for. You get the words and the attribution together, free on every transcript. For a two-person interview or a small meeting, that separation is usually accurate and saves the tedious work of tagging every line by hand. There is no cap on how many speakers a file can hold, though accuracy is highest when voices are distinct, do not overlap, and do not sound alike. That speaker labelling comes with every transcript at no extra cost is the kind of honest-by-default detail this whole approach is built on: the useful part is included, not held back behind a higher tier.

The plain distinction to hold on to is this. If you want to capture speech as you speak, voice typing or dictation, reach for the live tool built into your device or word processor. If you already have a recording you need as text, transcribing it from the audio-to-text page is faster, more accurate, and hands you speaker labels and exportable output in a single step.