Transcribe audio or video to text, or make subtitles (SRT/VTT). Free and unlimited; your file never leaves your device.
Built & maintained by Pappu Venkata Subbi Reddy, founder of Clacify · Updated October 2026
Audio & Video to Text turns speech into text without sending your recording anywhere. Drop in a meeting, lecture, interview, voice note, phone call or video, choose the language (or let it detect it), and you get a transcript with a timestamp on every line. Click a timestamp to hear that moment, fix any word in place, then copy the text or download it as a plain text file or as SRT/VTT subtitles for YouTube, VLC or a video editor. It can also translate speech in another language straight into English text. Most free transcription sites cap you at a few minutes or files a day, because every minute they transcribe costs them server time. This tool runs OpenAI's Whisper model on your own device, so there is nothing to meter: no minute limits, no sign-up, and no upload.
The file is decoded by your browser's own media decoder directly at 16 kHz, the sample rate Whisper was trained on, and mixed to mono. Decoding at 16 kHz rather than the usual 48 kHz keeps an hour of audio at about 230 MB of memory instead of well over a gigabyte, which is what makes long recordings practical in a browser tab. Recognition is done by Whisper, the open-source speech model OpenAI trained on 680,000 hours of multilingual audio, run through ONNX Runtime in WebAssembly inside a background worker so the page stays responsive. Audio is processed in 30-second windows that overlap by 5 seconds on each side; the overlaps are used to stitch the windows together so words at the boundaries are not cut or repeated, and every piece of text keeps its start and end time. The text appears on screen as each window is decoded, so you can read along instead of waiting for the end. You choose between three model sizes. Fast uses Whisper tiny (a 42 MB download), Balanced uses Whisper base (76 MB) and Best uses Whisper small (240 MB). The model downloads once from Hugging Face and is then cached by your browser, so later transcriptions start immediately. We use 8-bit weights because they are the smallest download. In our tests they ran faster on the processor than on the graphics card: the graphics path in browsers lacks fast 8-bit maths, and Whisper base took 20.4 seconds there against 8.9 seconds on the processor for the same 20-second clip. Choosing an Indian language switches to Best automatically, because the smaller models are noticeably weaker outside English. For "Translate to English", Whisper's built-in translation task is used: it listens to the speech in the source language and writes English text directly, rather than transcribing first and translating the text afterwards.
| Setting | Model | One-time download | Speed in our test* | Best for |
|---|---|---|---|---|
| Fast | Whisper tiny | 42 MB | 20 s of speech in 4.4 s | Quick drafts, older phones, clear English |
| Balanced | Whisper base | 76 MB | 20 s of speech in 8.9 s; 64 s in 19 s | Most English recordings |
| Best | Whisper small | 240 MB | 20 s of speech in 11.9 s | Indian languages, accents, noisy recordings |
*Measured on a laptop processor in Chrome, October 2026. Phones are slower; the first run also includes the model download.
On a clear English recording (a 64-second voicemail-style message), Balanced transcribed every ordinary word correctly and put each sentence on its own timestamped line. The mistakes were all names and acronyms: "Clacify" came out as "Clasify" or "classified", and "EMI" as "Emmy" or "ME". This is typical of speech models, which guess unfamiliar words from how they sound, and it is why every line is editable before you download. When the same speech had fan-like background noise mixed in at a level close to the voice, more errors crept in ("rupees" became "Groupies"). For noisy recordings, run the file through the Audio Enhancer first and transcribe the cleaned version.
A transcription service that runs on its own servers pays for every minute it processes, so its free plan has to stop somewhere: a few minutes per file, a few files a day, or a sign-up wall. Here the processing happens on your computer or phone, so a longer recording costs us nothing. The practical limits are your device's: up to 60 minutes per file (split longer recordings with the MP3 Cutter) and a speed that depends on your processor. The one thing that does use the internet is the first download of the model, which is the same open-source file for everyone and contains nothing about you or your recording.
Whisper understands Hindi, Telugu, Tamil, Kannada, Malayalam, Marathi, Bengali, Gujarati, Punjabi and Urdu, but its accuracy in these languages is lower than in English, because far less training audio existed for them. Use the Best setting (selected automatically when you pick one of these languages), set the language yourself rather than relying on detection, and expect to correct some words. Speech that switches between Hindi and English mid-sentence is the hardest case. If what you need is English text from Hindi or Telugu audio, try "Translate to English": it often reads more naturally than a word-for-word transcript.
Drop the video file in directly; the tool takes the audio track itself. After transcribing, download SRT for YouTube, VLC, Premiere Pro or DaVinci Resolve, or VTT for web players. Each subtitle keeps the start and end time Whisper measured for that sentence. Read through the transcript before downloading: correcting a name once in the editor fixes it in every file you download.
Open Clacify's Audio & Video to Text, drop in your file (MP3, M4A, WAV, OGG, FLAC, or an MP4/WebM video) and press Transcribe. The first time, your browser downloads the speech model (42–240 MB depending on the accuracy you pick); after that it is cached. The transcript appears with timestamps, and you can copy it or download it as TXT, SRT or VTT.
Yes. Other free tools run the speech model on their own servers, which costs them money per minute, so they cap usage or ask you to sign up. Here the model runs on your own device, so there is nothing to meter. The practical limit is 60 minutes per file at a time, set by browser memory; split longer recordings with the MP3 Cutter.
Yes. Choose the language, and use the Best accuracy setting, which the tool selects automatically for Indian languages; the smaller models are noticeably weaker on them. Accuracy for Indian languages is lower than for English, especially with background noise or mixed Hindi-English speech, so check the result before relying on it.
Yes. Drop in the video file and, after transcribing, download the SRT or VTT file. SRT works with YouTube, VLC and most video editors; VTT is the web standard. Each line keeps its start and end time, and any corrections you make in the transcript are included in the download.
No. The audio is decoded and transcribed inside your browser tab. The only network request is the one-time download of the open-source Whisper model from Hugging Face, which contains no information about you or your file.