How to transcribe audio to text online in 3 steps
- 01
Open your recording or record with your microphone
Use Open file, drop your audio or video here, or record directly with your microphone. Speech to text also reads the audio track out of video files your browser can decode.
- 02
Choose a language, task, and model size
Let language auto-detect run, or pick the spoken language yourself. Choose transcribe to keep the original language or translate to get English text back, then pick a Whisper model size to balance speed and accuracy.
- 03
Edit the transcript, then export
Read through the timestamped transcript and fix anything the model missed. When it's ready, export it as TXT for a plain document, or SRT or VTT for subtitles.
What you can do with this speech to text tool
- Transcribe on-device with OpenAI's Whisper models - Tiny, Base, and Small (multilingual), plus Tiny.en and Base.en (English-only) - with no audio sent to a server.
- Auto-detect the spoken language across the many languages Whisper supports, or set it manually if you already know it.
- Translate directly to English text instead of transcribing in the original language, using the same Whisper model.
- Get timestamped segments for every line, so you can jump to the exact moment a phrase was spoken.
- Edit the transcript by hand before exporting, to fix names, technical terms, or anything the model heard imperfectly.
- Export as plain TXT for a document, or SRT and VTT for subtitle files ready to attach to a video.
- Pull audio straight out of video files your browser can decode, so you can transcribe a video without extracting the audio track yourself first.
What is speech to text online?
Speech to text online is the process of turning spoken audio into written text inside your browser, without sending the recording anywhere. Audio Lab's speech to text tool runs OpenAI's Whisper models on-device: the model downloads once and is cached by your browser, and every transcription after that runs locally on your machine.
You can transcribe a file you already have, record fresh audio with your microphone, or pull the audio track out of a video file. The output is a timestamped, editable transcript you can read through, correct, and export in the format your project needs.
Choosing a Whisper model size
Whisper comes in several sizes, and the tool lets you pick which one runs. Smaller models start faster and transcribe more quickly; larger models take longer to load and run but tend to pick up more nuance in the audio. There's a genuine tradeoff here rather than one obviously correct setting - it depends on how long your recording is and how much time you want to spend waiting.
| Model | Language support | Typical use |
|---|---|---|
| Tiny / Tiny.en | Multilingual or English-only | Fast drafts, short clips, quick checks |
| Base / Base.en | Multilingual or English-only | A balance of speed and detail for everyday use |
| Small | Multilingual | Longer or less clear recordings where extra care helps |
How accurate is speech to text transcription?
Accuracy depends on the recording, not just the model. Clear speech, one speaker at a time, and minimal background noise give any of the Whisper models a strong chance of getting names, numbers, and technical terms right. Overlapping speakers, heavy accents the model hasn't seen much of, music underneath dialogue, or a noisy room all make the job harder.
Model size matters too: a larger model generally handles difficult audio more gracefully than a smaller one, at the cost of taking longer to run. If a transcript comes back with obvious mistakes, cleaning up the recording first - running it through the noise remover - or choosing a larger model are both worth trying before you give up on a clip.
SRT vs. VTT: which subtitle format do you need?
Both formats carry the same information - timestamped lines of text - but they aren't interchangeable everywhere. Export whichever your video editor, hosting platform, or player expects; if you're not sure, SRT is the more widely recognized default.
| Format | Typical use | Notes |
|---|---|---|
| SRT | Widely supported subtitle format for video editors and players | Plain text with numbered blocks and timestamps |
| VTT | Web video subtitles, used natively by HTML5 video players | Similar to SRT with a web-friendly header and styling hooks |
| TXT | A plain transcript with no timing | Well suited to show notes, articles, or reading rather than captioning |
Speech to text vs. text to speech vs. voice cloning
These three tools sit in opposite directions of the same idea, and it's easy to mix them up by name. Speech to text goes from audio to words - it's the one you want for transcripts and subtitles.
Text to speech goes the other way, turning written text into narration using a built-in voice. Voice cloning also generates audio, but in a specific reference voice you provide rather than a fixed library. If you're starting from a recording and need words on a page, this is your tool.
Limitations and what to expect
Transcription length is device dependent, with a practical ceiling around 60 minutes per recording - longer files need more memory and more time to process, and how far you can push it depends on your device. Multiple overlapping speakers are transcribed as well as the model can manage, but speech to text doesn't separate or label who said what.
Whisper's language auto-detect covers a very wide range of languages, but detection can be shaky on very short clips or on languages it has seen less training data for - if you know the language, setting it manually is more reliable than auto-detect. Background music playing under dialogue, cross-talk between multiple people, and strong regional accents are the situations most likely to need a manual re-check of the transcript afterward.
Tips for a cleaner transcript
- Record close to the microphone and away from fans, traffic, or music playing in the background - clean audio in is the single biggest factor in transcript quality.
- Try a smaller model first for a quick check, then re-run a longer or harder recording on a bigger model if the draft has too many mistakes.
- Set the language manually instead of relying on auto-detect for very short clips, where there isn't much audio for detection to work with.
- Run noisy audio through the noise remover before transcribing if hiss, hum, or fan noise is making the model guess.
- Read the transcript against the audio once before exporting - names, brand terms, and numbers are the most common things worth double-checking by hand.
- Choose translate instead of transcribe if you need English text from a foreign-language recording, rather than transcribing and translating separately.
Frequently asked questions
Is this speech to text tool free?
Yes. There's no sign-up and no limit on how many recordings you transcribe. The Whisper model you choose downloads once and is cached by your browser, so later transcriptions don't need to fetch it again.
Does my audio get uploaded to a server?
No. Transcription runs on-device using the Whisper model your browser downloaded and cached. Your recording stays on your machine throughout, which makes this a reasonable option for interviews, meetings, or anything you'd rather not send anywhere.
How accurate is the transcription?
It depends on your audio and the model size you pick. Clear, single-speaker recordings with little background noise transcribe more reliably than noisy or overlapping audio. Larger Whisper models tend to handle difficult audio better than smaller ones, at the cost of taking longer to run - there's no single accuracy number that applies to every recording.
What languages are supported?
The multilingual Whisper models (Tiny, Base, Small) support a wide range of languages with auto-detect or manual selection. The English-only variants, Tiny.en and Base.en, are tuned specifically for English and can be faster for English audio.
How long can a recording be?
Practically, up to around 60 minutes, and the real limit depends on your device's memory and processing power rather than a fixed number. Very long files take proportionally longer to transcribe and use more memory while processing.
Can I transcribe a video file?
Yes. Speech to text reads the audio track out of video files your browser can decode, so you can drop in a video the same way you would an audio file and get a transcript back.
Can I get subtitles instead of a plain transcript?
Yes. Export the timestamped transcript as SRT or VTT for subtitle files, or as TXT if you just want a plain document without timing information.
Why does the first transcription take a while?
The first time you use a given Whisper model, your browser has to download and cache it before transcription can start. Once it's cached, later transcriptions with that model size start much faster.
Does this work on mobile, including iPhone?
It runs in a mobile browser, but Whisper models use meaningful memory and processing power, so a phone may run slower or run into memory limits on longer recordings. A desktop or laptop browser handles longer files more comfortably.
Can I edit the transcript before exporting?
Yes. The transcript is editable directly in the tool, so you can fix misheard names, technical terms, or punctuation before you export TXT, SRT, or VTT.
What is the difference between transcribe and translate?
Transcribe keeps the output in the spoken language of the recording. Translate uses the same Whisper model to produce English text instead, regardless of the original language - useful when you need an English transcript from foreign-language audio.
How is this different from text to speech?
This tool goes from audio to text. Text to speech does the opposite, turning written text into narration with a built-in voice. If you're starting from a recording, speech to text is what you want.