Speech to text
Dictate, or drop in a voice memo or a call recording and get the text back by speaker. It runs in your browser, so the audio never leaves your device.
- Free, no sign-up
- 10 languages
- No audio uploaded
The English model (71 MB) downloads the first time you start.
The audio stays on this device. The model downloads once and stays in your browser for next time.
Speech model: Kroko ASR community model by Banafo, CC BY-SA 4.0.
How it works
Pick a language
The first time, your browser downloads the speech model for it, 29 to 233 MB. After that it starts from your browser's storage.
Talk or add a file
Dictate into your microphone, or drop in a voice memo, a voicemail or a recording: MP3, M4A, WAV, OGG or WebM. A call or a meeting comes back by speaker.
Take the text
Copy it, save it as a text file, or turn it into call notes with the call notes generator.
Why it runs on your device
Most online transcription uploads your audio to a server. This tool does the recognition inside your browser, with open speech models. Dictation uses the same ones as Superscribe Lite on Mac and Windows. Your recording is never sent anywhere, so you can use it for calls and memos you would not hand to a stranger's server.
The models are Kroko by Banafo for English, German, French, Spanish, Italian, Dutch, Portuguese and Swedish (CC BY-SA 4.0), the Estonian model from Tallinn University of Technology (MIT), and Vosk for Russian (Apache 2.0). Audio files in English and German go to NVIDIA's Parakeet and FastConformer (CC BY 4.0) and Russian files to GigaAM v3 by Sber (MIT): they read each stretch of speech whole rather than word by word as you talk, and make fewer mistakes. Silero VAD (MIT) finds where the speech is. The speaker labels come from pyannote (MIT) and 3D-Speaker (Apache 2.0).
More free tools
- Call notes generatorPaste a call transcript and get the note: a summary, the follow-ups and an entry for your CRM.
- Attendance note generatorTurn a client call into the attendance note for the file: what the client said, the advice given, the next steps and the time spent.
- Candidate call notesTurn a screening call into notes on the candidate: fit for the role, pay, notice period and a summary for your client.
- Viewing feedback generatorTurn a buyer's call after a viewing into the feedback note, and a short update you can send the seller.
- Call recording lawsCan you record this call? A plain answer for every US state and 33 countries, with the law behind it.
Questions
Is it really private?
Yes. The page downloads a speech model and runs it in your browser. Your audio is not uploaded to Superscribe or to anyone else, and you can check that in your browser's network panel while you dictate. Text leaves the page only when you press Find names: then Claude, the AI from Anthropic, reads the transcript for the speakers' names, and Superscribe does not save it.
Which languages can it transcribe?
English, Estonian, German, French, Spanish, Italian, Dutch, Portuguese, Swedish and Russian. Pick the language before you start: each one has its own model.
Does it add punctuation and capital letters?
Yes, in every language but Russian. The models for English, Estonian, German, French, Spanish, Italian, Dutch, Portuguese and Swedish write capital letters, commas, question marks and sentence endings as you speak, though they miss one now and then, most often at the very end. The Russian model writes lowercase words without punctuation.
Does it tell who said what?
Yes, in audio files. When a recording has more than one voice, the text comes back as turns, Speaker 1, Speaker 2 and so on, and you can type in their names, or press Find names to fill in the names people say in the recording, such as "Hi, this is Anna". It is best on calls and meetings where people take turns; a word or two spoken over someone else can land on the wrong name. The first file also downloads the speaker model, 46 MB, once. The voices are compared only inside that one recording, on your device, and nothing about them is kept.
How long can a recording be?
Up to 30 minutes and 200 MB per file. Split a longer recording into parts and open them one at a time: each part's text is added to the end. Dictation has no time limit.
Why does the first start take a while?
The first time you use a language, your browser downloads its speech model, between 29 and 233 MB. Audio files in English, German and Russian use a second, more accurate model that downloads the first time you open one. They stay in your browser's storage, so the next time it starts in a few seconds.
Does it work on a phone?
Yes, on a recent phone: in Chrome, Edge or Firefox, or in Safari on iOS 16.4 or later. The larger models need more memory and processing power, so an older phone can be slow with them or run out of memory. English, French, German and Russian have the smallest models for dictation. Use Wi-Fi for the first download.
Dictate into any app, not just this page.
Superscribe Lite puts this on your Mac or PC: press a key, talk, and the text appears where your cursor is. It runs on your computer too.
Superscribe Lite is free, with no account.