How to transcribe a Zoom, Meet, or Teams recording with speaker labels

Transcribe the recording with speaker labels

Upload the file via POST /upload (up to 100 MiB, or extract the audio with ffmpeg first for longer recordings) or pass any public URL, then start the job. num_speakers is optional, though you should set it if you know how many people spoke.

curl https://app.kolwrite.com/transcribe \
  -H "x-api-key: $KOLWRITE_API_KEY" -H "Content-Type: application/json" \
  -d '{"audio_url": "https://example.com/weekly-sync.mp4", "model": "oracle", "language": "auto",
       "include_diarization": true, "num_speakers": 4, "output_format": "json"}'

Poll GET /transcribe/{jobId} for the result, or pass a webhook_url so we notify your server when a long recording is done.

Every segment includes the speaker label, timestamps, and word-level breakdown:

{
  "jobId": "34c63d45-4cb5-473b-bc60-25c175e69c57",
  "status": "COMPLETED",
  "result": {
    "language": "en",
    "segments": [
      {
        "start": 1.0,
        "end": 3.75,
        "text": "Hello everyone and welcome to the show.",
        "speaker": "Speaker 1",
        "words": [
          { "start": 1.0, "end": 1.35, "word": "Hello" },
          { "start": 1.4, "end": 1.75, "word": "everyone" }
        ]
      }
    ]
  }
}

Map speaker labels to real names

Speaker labels come back as generic strings (Speaker 1, Speaker 2). It is up to you to match them to real people, such as by checking the attendee list or seeing who introduces themselves first. A simple mapping in your code keeps it clear:

minutes.py

# minutes.py — turn the diarized result into text an LLM can summarize
NAMES = {"Speaker 1": "Dana", "Speaker 2": "Omar", "Speaker 3": "Lee"}  # you decide who is who


def clock(seconds):
    m, s = divmod(int(seconds), 60)
    h, m = divmod(m, 60)
    return f"{h}:{m:02}:{s:02}" if h else f"{m}:{s:02}"


def meeting_text(result):
    lines = []
    for segment in result["segments"]:
        who = NAMES.get(segment.get("speaker"), segment.get("speaker", "Unknown"))
        lines.append(f'[{clock(segment["start"])}] {who}: {segment["text"]}')
    return "\n".join(lines)


PROMPT = """Below is a meeting transcript with timestamps. Write: (1) a five-line summary,
(2) decisions, (3) action items as "owner - task - [timestamp]". Use only what is in the transcript.

"""
# Send PROMPT + meeting_text(result) to the model you use.

Minutes and action items

KolWrite gives you the timed, speaker-labeled transcript. For summaries and action items, pass that text to whichever LLM you already use. Leave the timestamps in the prompt so each action item can reference the exact moment in the recording, and instruct the model to stick strictly to what is in the transcript.

Call recordings with one speaker per channel

If you have a stereo call recording with the agent on the left channel and the customer on the right, you don't need diarization. Split the speakers by audio channel instead. Channel separation supports 2 to 8 channels and cannot be used alongside include_diarization.

{
  "audio_url": "https://example.com/support-call.wav",
  "model": "oracle",
  "channel_speaker_separation": "standard",
  "channel_labels": { "0": "Agent", "1": "Customer" },
  "output_format": "json"
}

Use overlap instead of standard when people talk over each other; it bills the model rate for the greater of the recording length or the total detected speech across channels.

Pricing

Standard diarization and standard channel separation don't cost anything extra. Full rates: /llms/pricing.md.

Things to keep in mind

• Meetings contain personal data. Make sure everyone on the call knows they are being recorded and transcribed, and secure the transcripts just as carefully as the audio files.
• Telling speakers apart is hardest when people talk over each other, have similar voices, or use poor microphones. For difficult audio, try pro_diarization or channel separation.
• If a meeting is in a single language, set language explicitly. Auto-detection only picks one language for the whole file.
• Have a human review names, numbers, and final decisions before anyone acts on the generated minutes.

Questions and answers