Opus is the codec almost every messenger uses for voice. WhatsApp stores each voice message as an .opus file, Signal records in the same codec, and Discord, Zoom and most browser-based call tools send Opus over the wire. It won that position because it stays intelligible at bitrates where MP3 turns to mush, which keeps a voice note small enough to send over a weak mobile connection.
The trouble starts when the file leaves the app. Share a WhatsApp voice message to your files or your email, and you hold an .opus that Windows Media Player, older versions of QuickTime and many audio editors will not open. People then look for an OPUS to MP3 converter when what they actually wanted was to know what the message says. This page skips that detour and goes from the file to the text.
On Android, WhatsApp voice notes sit in the phone storage under the WhatsApp media folder, in a directory called WhatsApp Voice Notes, sorted by week. On an iPhone, press and hold the message, choose Forward, then the share icon, and save it to Files. WhatsApp Web and the desktop app offer a Download option in the menu on each voice message. In an exported chat archive, the voice notes are the .opus files next to the text file.
Browser recorders are the other common source. A recording made with the MediaRecorder function of Chrome or Firefox is Opus audio, and tools built on it often save the result with an .opus or .webm ending. Discord recording bots and some VoIP systems write Opus too. An .opus file uses the OGG container underneath; the OGG to Text page explains how that container relates to .ogg and .oga files.
WhatsApp encodes voice notes in mono at a low bitrate, usually around 16 kbit/s. That sounds like very little, yet the codec was tuned for exactly this job and keeps the range where consonants live, so the speech model has what it needs. Converting the file to MP3 before uploading decodes it and compresses it a second time. The file gets bigger and the audio slightly worse, so upload the original.
The errors in a voice-note transcript nearly always come from the recording situation. People record while walking, driving or cooking, holding the phone at arm's length, and they speak informally, with names, slang and half sentences. Expect the occasional wrong word in those passages, and check names and numbers by ear. Each segment carries a timestamp, so the spot is easy to find in the audio.
Guests get 10 transcriptions, each up to 3 hours long, with 5 hours in total; the same allowance applies after you create a free account. Without an account you read and copy a free preview, which for a short voice note usually means the whole text. A free account holds the full transcript of longer recordings, with an editor and a history. Subtitle export as SRT, VTT or ASS, AI chat with the recording, summaries and PDF or Word export belong to the paid plans.
Voice messages are private by nature, and sometimes they are evidence: a verbal agreement, an instruction from a client, a message from a relative. The speech models run on servers we operate, and the audio is not forwarded to a third-party speech API.