How to use Transcribe audio and video
- Choose a recording: an MP3, M4A, WAV, OGG or FLAC file, or an MP4, MOV or WebM video.
- Leave the language on Detect automatically, or pick the language that is spoken.
- Select Transcribe. The first time, the speech model (about 50 MB) downloads once; then it works through the recording 30 seconds at a time on your device.
- Read the transcript, switch between plain text, SRT and WebVTT, and download or copy it.
Transcripts and subtitles without uploading the recording
Interviews, meetings, lectures, voice notes and videos are easier to search, quote and share as text. This tool listens to the recording and writes down what is said, with the time each line starts, so you get a readable transcript and subtitle files from the same run.
Transcription services usually upload the whole recording to their servers. This one runs a speech recognition model inside your browser instead. The file is decoded and transcribed on your device, so a confidential call, a patient interview or an unreleased video never leaves it, and WeaveAI never hears it.
How it works
The tool uses Whisper base, OpenAI's multilingual speech recognition model, released under the MIT licence, run by ONNX Runtime Web in a background worker in this tab. Your browser's own audio decoder first turns the file into sound at 16 kHz, so MP3, M4A, WAV, OGG, Opus and FLAC files work, as does the sound track of MP4, MOV and WebM videos.
Whisper listens to 30 seconds at a time. Each window becomes a spectrogram, the model writes the words it hears along with timestamps, and the next window starts where the last complete sentence ended, so words aren't cut in half at the boundaries. Long silences are recognised and skipped. The progress bar shows how much of the recording is done, and the latest line appears as it is written.
The language is detected from the first 30 seconds, or you can choose it from the list, which helps with short clips and strong accents. Whisper base handles 99 languages; widely spoken ones such as English, Spanish, French and German come out best.
Formats and limits
The plain text version joins the lines into paragraphs, starting a new one after a pause. SRT and WebVTT keep every line with its start and end time, ready for YouTube, video editors and the HTML video player. Copy the text or download any format.
On a recent laptop, a recording takes roughly as long to transcribe as it does to play, longer on older phones. Recordings up to an hour are accepted. Like any speech model, Whisper can mishear names, jargon and overlapping speakers, so read the transcript through before you publish it.
Questions
- Is my audio uploaded?
- No. The speech model is downloaded to your browser, and the recording is decoded and transcribed on your device. The only download is the model itself, from WeaveAI's own engine server; your audio never leaves your device.
- How big is the download, and how long does it take?
- The model is about 50 MB and downloads once, then your browser keeps it. Transcribing takes roughly as long as the recording on a recent laptop, and longer on older phones. An hour of audio is the limit.
- Which languages does it understand?
- Whisper base was trained on 99 languages and detects which one is spoken from the first 30 seconds. English, Spanish, French, German and other widely spoken languages work best; you can also pick the language yourself.
- Can I make subtitles for a video?
- Yes. Choose the video, then pick SRT or WebVTT: each line has the time it is spoken, ready for YouTube, video editors and HTML video players.