A Discord voice message is an ordinary attachment: an Ogg file with Opus audio, normally named voice-message.ogg. On a computer, in the desktop app or at discord.com in a browser, the message offers a way to download it, either as a download control on the audio player or through the message's context menu. If you only find an option to copy or open the attachment link, open that link in the browser and save the file from there.
On a phone it is less direct. The mobile apps are where voice messages are recorded, but they are built to play them, and a save option is not always there. If you cannot find one in the message menu, the dependable route is to open the same conversation on a computer, or in the phone's browser in desktop mode, and download the attachment there.
When you save several messages, rename them as you go. Every file arrives with the same name, and a folder of identical voice-message.ogg copies is hard to match back to who said what.
Voice messages are quick to send and slow to deal with. You cannot skim one, you cannot search a server for something that was only spoken, and you cannot listen at all in a meeting or with other people in the room. A few lines of text solve all three.
Community moderators use transcripts to document what was said in a report or an appeal. Study groups and project teams paste the text under the voice message so that the decision in it shows up in search later. People who are deaf or hard of hearing, and members who are not fluent in the language of the server, can read a message, or run it through a translator, when listening is not an option.
The second case is longer: a recorded voice channel. Tabletop role-playing groups record sessions to write recaps, podcast hosts record remote guests, and teams keep a record of a planning call. Discord does not record voice channels by itself, so this is done with a recording bot or with screen and audio capture software such as OBS.
Recording bots such as Craig can deliver the session in two ways: one mixed file with everybody in it, or a separate track for each speaker. For a transcript of the whole conversation, upload the mixed file. Upload individual tracks when you want one person's words only. Bear in mind that each track counts as a separate recording, so a six-person session uploaded track by track uses six of your 10 free transcriptions.
A two- or three-hour session fits within the 3 hours limit for one recording, and an hour of audio typically takes three to five minutes to process. Make sure everyone in the channel knows the session is being recorded; in many places that is a legal requirement and on any server it is good manners.
A voice message has one speaker, so there is nothing to tell apart. For a mixed channel recording it matters: the guest transcript on this page has timestamps but no speaker labels. Speaker labels are added to recordings you upload from inside a free account, where each voice the model can separate is marked as a distinct speaker. Per-speaker tracks from a bot sidestep the question, because each file is one person.
Usernames, game terms, character names and server slang are the words most likely to be misheard, so give them a quick check. People talking over each other, a keyboard next to the microphone and music bots playing in the channel all cost accuracy. The language is detected automatically from the audio, with more than fifty supported.
The speech models run on servers we operate, and your audio is not forwarded to an external speech API. Audio uploaded as a guest is deleted after a day, and the transcript stays so you can come back to it.
The free allowance is 5 hours in total across 10 recordings, shared between guest use and a free account. As a guest you get a preview, which shows a short voice message in full and the first part of a long session; the full transcript of a long recording is kept in a free account, with an editor and a history. AI summaries, AI chat, translation, subtitle export and PDF or Word export are part of the paid plans.