How to Transcribe a Discord Call — Free

Discord doesn't record voice channels natively. Here's the cleanest workflow: Craig bot for recording, ScribeForge for transcription. End-to-end, free.

Why Craig + ScribeForge beats single-file recording

The naive approach is to use OBS Studio or system audio recording — gives you one mixed audio file with all speakers overlapping. Grok STT can diarize this (Speaker 1, Speaker 2…) but accuracy drops on noisy multi-speaker mixes.

Craig records each speaker on a separate track. You get one FLAC per Discord username. Transcribe each FLAC individually → automatic per-speaker labeling, no diarization needed. Concatenate by timestamp at the end.

The 5-step workflow

1. Add Craig — visit craig.chat and invite the bot to your server. Admin permissions required. Free tier is enough for casual calls.
2. Start recording — in your voice channel, type :craig:, join. Craig joins as a participant and starts capturing. Announce the recording to all participants.
3. End recording — type :craig:, leave or end the call. Craig DMs the recording-host a download link.
4. Pick "Multiple FLAC tracks" — one file per speaker, lossless. Filename includes the Discord username, so you already know who is who.
5. Transcribe each track — drop each .flac into ScribeForge. Result is a clean per-speaker transcript. Combine by timestamp at the end.

Compress FLAC before upload (25 MB cap)

FLAC is lossless and large. A 1-hour call produces 30–100 MB per speaker — over the 25 MB cap. Convert to OGG Opus at 32 kbps mono first:

ffmpeg -i speaker.flac -c:a libopus -b:a 32k -ac 1 speaker.ogg

32 kbps Opus mono is plenty for transcription accuracy — the human voice fits in that bitrate cleanly. Result is 5–10× smaller, fits well under the cap.

Combining per-speaker transcripts

Each ScribeForge transcript includes phrase-level timestamps (start, end, text). For a unified call transcript:

1. Save each transcript as speaker-{username}.json (download includes timestamps).
2. Merge into a single timeline by start-time, prefixing each line with the speaker name.
3. One Python snippet does this:

import json, glob

events = []
for fn in glob.glob("speaker-*.json"):
    speaker = fn.replace("speaker-", "").replace(".json", "")
    for seg in json.load(open(fn))["segments"]:
        events.append((seg["start"], speaker, seg["text"]))

events.sort()
for t, who, text in events:
    print(f"[{t:7.2f}] {who}: {text}")

Common questions

OBS Studio captures system audio + your mic to a single file. Loses per-speaker separation, so diarization becomes Grok's job (works, just less clean). Browser-based discord clients can also be captured by browser extensions like "Voice Recorder" — same caveats.

Two-party-consent jurisdictions (most of EU, US states like California, Florida, Pennsylvania, …): all participants must consent. One-party-consent: only the recorder needs to know. Best practice always: announce the recording when Craig joins. ScribeForge does not record — only transcribes whatever audio you provide.

Discord ships voice over Opus — generally clean. On per-speaker tracks (Craig's multi-FLAC) accuracy is in the 5–8% WER range — close to studio quality. Mixed multi-speaker tracks degrade somewhat depending on overlap.

No — sent to xAI's Grok STT API for processing, deleted immediately. The transcript exists only in your browser session.

Related guides