Speech to text · Subtitles · Runs on your device

Audio to Text & Subtitle Generator

Open an audio or video file, generate a transcript with timestamps, fix any words or timings, then download TXT, SRT or VTT. Recognition runs in your browser with OpenAI's open Whisper model, so your recording is not uploaded.

No upload, no sign-upEdit text & timingTXT · SRT · VTT

1Open a file
2Generate
3Review & download

Open audio or video

or drop a file here · up to 500 MB and 30 minutes

Generate the transcript

What do you need? Transcribe Text in the language that is spoken Translate to English English text from speech in another language

Spoken language Accuracy Engine

Open a file to begin.

Review, correct and download

Your transcript will appear hereSegments show up while recognition runs. Click a time to jump there, edit any text, and fix timings with the “now” buttons while playing.

Automatic recognition makes mistakes, especially with names, numbers, accents, music, overlapping voices and background noise. Timestamps are approximate (often within about half a second) and are tidied using the audio's quiet parts. Segments marked check look suspicious, for example text that Whisper sometimes invents during silence. Please review before publishing.

How to create subtitles from audio or video

Choose an audio or video file (MP3, WAV, M4A, OGG, WebM, MP4 and others your browser can decode), or try the sample clip.
Pick Transcribe to keep the spoken language, or Translate to English. Set the spoken language if you know it; auto-detect works best on clear speech.
Press Generate transcript. The first time, the model downloads to your browser (about 41–240 MB depending on accuracy). Later runs reuse the saved copy.
Play the file and read along. Edit words directly, click a time to jump there, and use now to set a start or end to the current playback time. Split, merge, delete or add segments as needed.
Download SRT for most video editors and YouTube, VTT for web players, or TXT for a readable transcript.

Example

Open the sample clip and generate with Balanced. In our test it produced two segments, “Welcome to ToolSourceRour. Try a new voice, compare the sound, and make something creative.” and “This is a sample recording for the voice changer.” The model misheard the site name, so type “ToolSorcerer” into the first text box, then download SRT. The first block of the file reads 1, 00:00:00,000 --> 00:00:06,500 and the corrected text. With Most accurate the name was recognized correctly.

Questions & answers

Is my audio uploaded anywhere?

No. Decoding and recognition run in your browser. The only download is the speech model itself, which comes from Hugging Face the first time and is then stored in your browser's cache. Your audio and transcript are not sent to ToolSorcerer or any recognition service.

How big is the download and what device do I need?

Standard (CPU): Fast ≈ 41 MB, Balanced ≈ 76 MB, Most accurate ≈ 240 MB. The Graphics card (WebGPU) engine uses larger files (Fast ≈ 114 MB, Balanced ≈ 199 MB) and is offered for Fast and Balanced only. Use a recent desktop Chrome, Edge or Firefox with at least 4 GB of RAM; Most accurate needs more memory and can fail on phones or older computers.

How long does it take?

It depends on your device. In our tests on a desktop CPU, Fast processed audio at about 0.2× its length, Balanced at about 0.3–0.5× and Most accurate at about 0.9–1.4×. A 10-minute file on Balanced takes roughly 3–5 minutes there. Phones and older laptops are slower. WebGPU was about twice as fast as CPU for Balanced.

What is the difference between transcribe and translate?

Transcribe writes down what is said in the same language. Translate to English listens to speech in another language and writes English text. Whisper can only translate into English, and translation is less accurate than transcription. Choosing Translate for English speech simply gives an English transcript.

Which languages are supported?

Whisper was trained on about 99 languages. Quality varies a lot: widely spoken languages such as English, Spanish, French, German, Portuguese and Italian work best, and the smaller models are noticeably weaker in other languages. Our own tests used English speech only, so try a short clip first.

Why is a segment marked “check”?

The tool flags text that is often invented during silence or music (like “Thank you.”), repeated lines, text that is too fast for its time span, very long segments and estimated end times. A flag is a hint, not proof of an error.

Can I cancel or keep partial results?

Yes. Cancel stops recognition immediately. Segments finished so far stay in the editor and can be exported. Starting again replaces them.

Does it identify speakers or work live?

No. It does not label who is speaking, and it processes files rather than live microphone input.

Using Audio to Text & Subtitle Generator

Transcribe audio or video files in your browser with the open Whisper speech model, correct the words and timestamps, and download TXT, SRT or VTT. Translation to English is optional.

Choose an audio or video file, or use the sample clip.
Pick Transcribe or Translate to English, the spoken language and an accuracy level.
Press Generate transcript. The model downloads once (about 41–240 MB) and is cached in your browser.
Play the file, correct the text, and adjust start/end times with the now buttons, split, merge or delete.
Download SRT, VTT or TXT.

Example

Open the sample clip and generate with Balanced. In our test the first segment read “Welcome to ToolSourceRour. Try a new voice, compare the sound, and make something creative.”
Correct the name to “ToolSorcerer”, then download SRT; the first block starts 00:00:00,000 --> 00:00:06,500.

Questions & answers

Is my audio uploaded anywhere?

No. Decoding and recognition run in your browser. Only the speech model downloads from Hugging Face the first time and is cached. Your audio and transcript are not sent anywhere.

How big is the download and what device do I need?

Standard (CPU): Fast ≈ 41 MB, Balanced ≈ 76 MB, Most accurate ≈ 240 MB. The WebGPU engine uses larger files (Fast ≈ 114 MB, Balanced ≈ 199 MB). Use a recent desktop Chrome, Edge or Firefox with at least 4 GB of RAM.

How long does it take?

On a desktop CPU in our tests: Fast about 0.2× the audio length, Balanced 0.3–0.5×, Most accurate 0.9–1.4×. WebGPU was about twice as fast for Balanced. Phones and older laptops are slower.

What is the difference between transcribe and translate?

Transcribe keeps the spoken language. Translate to English writes English text from speech in another language and is less accurate. Whisper only translates into English.

Which languages are supported?

Whisper was trained on about 99 languages, with the best quality in widely spoken ones. Our own tests used English speech only, so try a short clip first.

Why is a segment marked check?

It looks like text Whisper invents during silence, repeats the previous line, is too fast for its time span, is very long or has an estimated end time. It is a hint, not proof of an error.

Does it identify speakers or work live?

No. It does not label speakers and processes files rather than live microphone input.