Whisper Tiny
About 45 MBThe quickest option for drafts and clear, simple recordings. Less reliable with background noise, accents and specialist vocabulary.
AUDIO TO WORDS, WITHOUT THE FRICTION
Every word.
In your language.
Turn recordings into readable transcripts and ready-to-use subtitles. All on your device.
Multilingual WhisperNo account neededNo audio uploads
01
Your recording
or drop a file here
MP3, WAV, M4A, FLAC, OGG or browser-supported video
No app-imposed size or duration limit
No app-imposed size or duration limit
Choose the language if you know it. Speech stays in its original
language.
Model files download on first use, then stay cached when your browser allows it. Keep this tab open while processing.
02
Your transcript
Ready when you are
Preparing recording
Preparing recording0s elapsed
The first run includes a model download.
A clear record of the conversation.
Choose a recording to get started.
Your transcript and
subtitle downloads will appear here.
Your audio stays on this device. Review names, numbers and important details before sharing a transcript.
Choose the right balance
Base is a good starting point. Small offers stronger multilingual accuracy. Tiny prioritizes speed. All three support Portuguese and many other languages.
Know what to expect
The percentage shows how much audio has been processed. It advances as speech segments finish, so pauses between updates are normal. Model preparation is shown separately.
Keep it yours
No recording is sent to a server. Download plain text for reading, or SRT subtitles for your video editor.
About the models
All three are multilingual Whisper models. They transcribe speech in its original language, including Portuguese, rather than translating it into English.
Whisper Base
About 80 MB · DefaultA balance of speed and accuracy for everyday recordings. Start here for interviews, voice notes and conversations.
Whisper Small
About 255 MBMore capacity for multilingual transcription and detailed speech. Requires more memory and processing time; accuracy still depends on the recording.
Model downloads are approximate and cached when possible. Processing uses compressed 8-bit models on your device’s CPU; the shared runtime adds about 22 MB on first use. Clear audio and selecting the spoken language help. These models do not identify speakers, and important details should be reviewed. Long recordings need more memory; practical limits depend on your browser and device.