IETF Meeting vCons

This repository contains vCon (Virtual Conversation Container) files for IETF working group sessions: the numbered meetings 66-126 (July 2006 - July 2026), 61 consecutive meetings and 8,179 sessions, plus the VCON working group's two 2026 interim sessions.

What is vCon?

vCon is an IETF standard format for capturing conversation data. Each vCon file contains:

• Meeting metadata - Date, location, working group information
• Video recording - YouTube URL for the session recording
• Transcript - Full transcript in WTF (World Transcription Format) with word-level timestamps
• Materials - Links to slides, agenda, minutes, and other session documents
• Participants - The working group chairs who were serving at the time of the session, plus an attendees party
• Lawful basis - IETF Note Well documentation per draft-howe-vcon-lawful-basis

Repository Structure

One directory per meeting, holding one vCon per working group session. A numbered meeting is ietf<N>/; an interim is named as the Datatracker names it, since an interim has no number:

ietf-meeting-vcons/
├── ietf66/                    # IETF 66 (July 2006, Montreal)
│   ├── ietf66_dnsop_1234.vcon.json
│   └── ...
├── ietf67/
├── ...
├── ietf126/                   # IETF 126 (July 2026, Vienna)
├── interim-2026-vcon-01/      # VCON interim, 2026-01-27
└── interim-2026-vcon-02/      # VCON interim, 2026-09-09

Interim sessions

Interim coverage is not a backfill. Working groups have held hundreds of interims and only the VCON working group's are here, generated the same way and from the same Datatracker documents as everything else. Any loader that walks vcon_dir recursively picks them up without changes; one that assumes an ietf<N> directory name will miss them.

The dataset spans four eras, which differ in what the IETF published at the time rather than in how the vCons are built:

Meetings	Period	What exists	Dialog
66-89	2006-2014	Agendas, minutes, slides. No recordings.	none
90-94	2014-2015	Audio MP3 on ietf.org, some YouTube video	audio/mpeg or video/mp4
95-125	2016-2026	YouTube video, most with auto-captions	video/mp4
126	July 2026	YouTube video with auto-captions	video/mp4

A vCon is not required to have a dialog. The 66-89 meetings have no recordings, but every one of those sessions carries its agenda, minutes and slides, which is still a record of the session worth having.

Transcript bodies for meetings 66-125 are not in this repository. They are published as GitHub Release assets and referenced from the vCons by URL and hash. See Transcripts below.

Getting the data

Git — the whole dataset with history:

git clone https://github.com/vcon-dev/ietf-meeting-vcons.git

S3 — anonymous read, no AWS account or credentials required. Useful for pulling a single meeting, or for streaming the set without a 140 MB clone:

s3://ietf-meeting-vcons/{meeting}/{filename}.vcon.json
https://ietf-meeting-vcons.s3.amazonaws.com/{meeting}/{filename}.vcon.json

# one session, over plain HTTPS
curl -O "https://ietf-meeting-vcons.s3.amazonaws.com/ietf126/ietf126_vcon_35521.vcon.json"

# one meeting
aws s3 sync s3://ietf-meeting-vcons/ietf126/ ./ietf126 --no-sign-request

# browse or mirror the whole set (--no-sign-request = no credentials)
aws s3 ls s3://ietf-meeting-vcons/ --recursive --no-sign-request
aws s3 sync s3://ietf-meeting-vcons/ ./ietf-vcons --no-sign-request

The bucket is read-only to the public: s3:GetObject and s3:ListBucket are granted anonymously, writes are denied. It carries all 8,181 vCons and nothing else — transcript bodies stay on GitHub Releases, which is where the vCons reference them.

Deploying this dataset

dataset.json at the repo root declares the whole repository as a single dataset, so a loader can ingest it without being told how the files are laid out:

{
  "name": "ietf-meeting-vcons",
  "version": "2.1.0",
  "vcon_dir": ".",
  "count": 8181,
  "spec": "0.4.0"
}

vcon_dir is the root, and vCons live in per-meeting subdirectories, so a loader should walk it recursively. count is a checksum on completeness: a partial or truncated checkout can be rejected instead of silently ingesting fewer sessions than the dataset claims.

File Naming Convention

Files follow the pattern: {meeting}_{group}_{session_id}.vcon.json

• meeting - ietf plus the meeting number for a numbered meeting (ietf124), or the Datatracker's interim name for an interim (interim-2026-vcon-02)
• group - Working group acronym (e.g., httpbis, quic, tls)
• session_id - Unique session identifier from the IETF Datatracker

vCon Structure

Each vCon file follows the draft-ietf-vcon-vcon-core specification:

{
  "vcon": "0.4.0",
  "uuid": "019d3273-6d8e-877c-9dd8-dd37220d739c",
  "created_at": "2026-07-22T12:00:00+00:00",
  "subject": "IETF 126 - VCON Working Group Session",
  "extensions": ["lawful_basis", "role", "meta", "wtf_transcription"],
  "parties": [
    {"name": "Chair Name", "mailto": "chair@example.com", "role": "chair"}
  ],
  "dialog": [
    {"type": "recording", "mediatype": "video/mp4", "url": "https://youtu.be/..."}
  ],
  "attachments": [
    {"purpose": "agenda", "url": "https://www.ietf.org/proceedings/126/agenda/agenda-126-vcon-00.md",
     "content_hash": "sha512-...", "mediatype": "text/markdown", "filename": "agenda-126-vcon-00.md",
     "meta": {"title": "VCON Agenda", "datatracker_url": "https://datatracker.ietf.org/meeting/126/materials/agenda-126-vcon"},
     "party": 0, "dialog": 0},
    {"purpose": "slides", "url": "https://www.ietf.org/proceedings/126/slides/slides-126-vcon-chair-slides-00.pdf",
     "content_hash": "sha512-...", "mediatype": "application/pdf", "party": 0, "dialog": 0},
    {"purpose": "lawful_basis", "encoding": "json", "body": "{\"lawful_basis\": \"legitimate_interests\", ...}"}
  ],
  "analysis": [
    {"type": "wtf_transcription", "dialog": 0, "vendor": "youtube", "encoding": "json", "body": "{...}"}
  ]
}

Note the details that follow from draft-ietf-vcon-vcon-core-02, the revision this dataset targets (the draft is now at -03): the syntax parameter is 0.4.0, recordings use dialog type recording rather than video, attachments use purpose rather than type, and every body is a string rather than an object, with encoding declaring how to read it.

Why materials point at www.ietf.org rather than the Datatracker

A material attachment carries url + content_hash, so the URL has to return the hashed bytes. The Datatracker's /meeting/N/materials/<doc> is a display endpoint, not a file: it renders Markdown agendas as HTML pages and converts PowerPoint decks to PDF, so a hash of the published file cannot describe what it serves. https://www.ietf.org/proceedings/{meeting}/{type}/{filename} serves each file verbatim, is the tree rsync.ietf.org::proceedings mirrors, and is reachable by scripted clients across all eras. The Datatracker page is still the right thing for a human to open, so it is kept in meta.datatracker_url.

Landing pages -- the session page, the collaborative notes, a draft's own page -- are mutable HTML and carry no hash. They stay body references, since a content_hash on them would assert an integrity guarantee that does not hold.

Transcripts

There are 4,059 transcripts in WTF format, covering every session with a recording except the gaps noted under Coverage gaps. They are stored in one of two ways.

IETF 126 is inline. The transcript body sits in the vCon, so each file is self-contained and needs no network access to read. This is the simplest form and works well for a single meeting.

IETF 66-125 are externally referenced. A WTF body runs 200-400 KB, so inlining all of them would make this repository well over a gigabyte. The core spec anticipates this: the Analysis Object section states that it "SHOULD contain the body and encoding parameters or the url and content_hash parameters." Those meetings therefore reference their transcripts instead of carrying them:

{
  "type": "wtf_transcription",
  "dialog": 0,
  "vendor": "youtube",
  "product": "auto-captions",
  "url": "https://github.com/vcon-dev/ietf-meeting-vcons/releases/download/transcripts-ietf126/ietf126_vcon_35521.wtf.json",
  "content_hash": "sha512-09FC_f0yKzzG1V6tih0kmrJm2Fd4HBHJPd7zqYXe4aiZcmnj0-1XEbReZmty25QCkIDhpgxOAxAkGA773ORZjA",
  "mediatype": "application/json"
}

This keeps the repository around 120 MB while preserving integrity: the content_hash is a SHA-512 digest of the exact bytes served, formatted as sha512- plus unpadded base64url, and it is covered by the vCon's own signature. A substituted transcript fails verification exactly as an edited inline one would.

Reading a transcript either way

import base64, hashlib, json, urllib.request

def load_transcript(analysis):
    """Return the WTF body whether it is inline or externally referenced."""
    if "body" in analysis:
        return json.loads(analysis["body"])

    raw = urllib.request.urlopen(analysis["url"]).read()

    algorithm, _, expected = analysis["content_hash"].partition("-")
    digest = hashlib.new(algorithm, raw).digest()
    actual = base64.urlsafe_b64encode(digest).rstrip(b"=").decode()
    if actual != expected:
        raise ValueError(f"content_hash mismatch for {analysis['url']}")

    return json.loads(raw)


with open("ietf126/ietf126_vcon_35521.vcon.json") as f:
    vcon = json.load(f)

for analysis in vcon.get("analysis", []):
    if analysis["type"] == "wtf_transcription":
        wtf = load_transcript(analysis)
        for segment in wtf["segments"][:5]:
            print(f"[{segment['start']:.1f}s] {segment['text']}")

Always check the hash before trusting a fetched body. That check is the only thing separating an external reference from an unverified download.

Regenerating the split

# Externalize a meeting's transcripts (writes bodies to transcripts/ietf<N>/)
python scripts/externalize_transcripts.py --meeting 125 \
    --url-prefix https://github.com/vcon-dev/ietf-meeting-vcons/releases/download/transcripts-ietf125 \
    --body-dir transcripts/ietf125

# Confirm every reference resolves against its local body file
python scripts/externalize_transcripts.py --meeting 125 \
    --body-dir transcripts/ietf125 --verify

transcripts/ is gitignored. Publishing a meeting means uploading that directory's .wtf.json files as assets on a release tagged transcripts-ietf<N>, matching the URL prefix baked into the vCons.

Chairs

Chair parties name the people who chaired the session on the date it was held, not today's chairs. This matters: ietf2vcon reads the Datatracker's current group record, which would otherwise stamp 2026 chairs on a 2006 session. scripts/fix_historical_chairs.py resolves them per date from grouphistory and rolehistory.

The Datatracker's role history begins 2011-12-09, when the feature was switched on, so sessions before IETF 83 (March 2012) cannot be resolved. Rather than assert today's chairs on a session from 2006, those vCons list no chair party at all and keep only the attendees party. 2,006 vCons are in that state. Absence of a claim is preferable to a false one; the chairs are recoverable from the IETF proceedings pages for anyone who wants to do that work.

Data Sources

All data is sourced from public IETF resources:

• Session metadata: IETF Datatracker API
• Video recordings: IETF YouTube Channel
• Transcripts: three provenances, distinguished by vendor and able to coexist in analysis:
  • youtube — YouTube auto-generated captions, the default where they exist.
  • whisper — local Whisper (IETF 125), higher quality with word-level timing.
  • apple (product: speechanalyzer) — Apple's on-device Speech framework via the mi CLI, used where no captions exist: the audio era (IETF 90-94) and the caption-less 2016-2017 videos (IETF 95-98). Segment-level timing, no per-word timestamps or confidence.
• Materials: IETF Meeting Materials Archive

Media references

content_hash on an external dialog/attachment/analysis reference lets a consumer verify the bytes it fetches match what was hashed at dataset-build time. Two categories of reference in this dataset don't carry one, and won't:

• YouTube recordings (3,654 sessions, IETF 95-126). A YouTube URL serves an HTML page, not the media file, so fetching it and hashing the response would hash a web page, not a recording. No content_hash is written for these, and none should be.
• IETF 90-94 audio (425 sessions). The ietf.org/audio/... MP3 URLs moved behind IETF SSO login in 2026 (see Coverage gaps); a plain, polite fetch — no browser spoofing, no bot-protection evasion — gets redirected into the login flow rather than the file. The Datatracker's own recording-* documents for these sessions point at the same now-gated URL, there is no alternate public mirror, and the Meetecho recording-playback system only covers IETF 98 onward. These URLs are left as-is, unhashed, until IETF publishes a public replacement.

Every other external reference in the dataset (slides, agendas, minutes, transcript bodies) is hashed.

IETF Note Well

All IETF meeting sessions are conducted under the IETF Note Well, which permits recording, transcription, and publication. This is documented in each vCon's lawful_basis attachment.

Statistics

Metric	Value
Meetings	61 numbered (IETF 66-126, consecutive) plus 2 VCON interims
Total vCons	8,181
Sessions with a recording	4,079 (3,654 video, 425 audio)
Transcripts	4,061 (142 inline, 3,919 externally referenced)
Transcript coverage	4,061 of 4,079 recordings (99.6%)
Date Range	July 2006 - September 2026
Working Groups	~50 per meeting
Repository size	~140 MB (plus ~1.1 GB of transcript bodies published separately)

Coverage gaps

18 of 4,079 recorded sessions have no transcript. Nearly all are YouTube videos that have since been removed or made private, so there is no longer any source to transcribe; a few are recordings that are silent or non-speech.

Everything else with a recording is transcribed. The audio era (IETF 90-94), whose recordings moved behind IETF SSO in 2026, was fetched through an authenticated session and transcribed on-device (see Transcribing without captions); the caption-less 2016-2017 videos (IETF 95-98) the same way.

IETF 66-89 has no recordings by design. The IETF did not publish session recordings in that era, so those 2,887 vCons are materials-only. They are not empty: they carry 20,855 slide decks, 5,721 agendas and 5,634 minutes, a median of 9 documents per session.

ietf107 (March 2020, the first virtual meeting) is a special case within the recorded era: it has 129 vCons and no recordings at all, because the Datatracker holds no recording documents for it. Its sessions are materials-only. A vCon is not required to have a dialog; the agenda, slides and minutes still make these a useful record of the session.

Usage Examples

Python

import json

# Load a vCon
with open("ietf121/ietf121_quic_33502.vcon.json") as f:
    vcon = json.load(f)

# Get session info
print(f"Subject: {vcon['subject']}")
print(f"Recording: {vcon['dialog'][0]['url']}")

# Chairs
for party in vcon["parties"]:
    if party.get("role") == "chair":
        print(f"Chair: {party['name']}")

For transcripts see Reading a transcript either way; meetings 66-125 reference them by URL rather than carrying them inline.

jq (Command Line)

# Get all recording URLs from a meeting
jq -r '.dialog[0].url // empty' ietf121/*.vcon.json

# Extract transcript text (IETF 126, inline bodies; note body is a JSON string)
jq -r '.analysis[] | select(.type=="wtf_transcription") | .body | fromjson | .segments[].text' \
    ietf126/ietf126_vcon_35521.vcon.json

# List transcript URLs for an externally referenced meeting
jq -r '.analysis[]? | select(.type=="wtf_transcription") | .url' ietf121/*.vcon.json

# List all working groups in a meeting
ls ietf121/*.vcon.json | sed 's/.*ietf121_\(.*\)_.*/\1/' | sort -u

Generation Tool

These vCons were generated using ietf2vcon, an open-source tool for converting IETF meeting sessions to vCon format.

To generate additional vCons:

pip install ietf2vcon
ietf2vcon convert --meeting 126 --group quic

Adding a New Meeting

A whole meeting is converted, normalised and validated in three steps. IETF 126 was produced this way:

ietf2vcon convert-all --meeting 126 --output-dir ./build126 \
    --transcript-source youtube --parallel 4

ietf2vcon stores WTF transcripts as an attachment, while this repository (and draft-howe-vcon-wtf-extension, which defines WTF as an analysis type) keeps them in analysis. After copying the *.vcon.json files into ietf126/, normalise and bring them into draft-ietf-vcon-vcon-core-02 compliance:

python scripts/wtf_attachment_to_analysis.py --meeting 126
python scripts/migrate_compliance_core02.py --meeting 126

Both scripts accept --dry-run and -v, and both are idempotent: re-running them on an already-migrated meeting is a no-op.

Backfilling Transcripts

If sessions end up with a recording but no transcript, fetch the YouTube auto-captions directly rather than regenerating the vCons, which would churn their UUIDs:

python scripts/backfill_youtube_captions.py --meetings 110-124 --workers 2 --delay 2

It skips sessions that already have a YouTube transcript, so it is idempotent and resumes cleanly after an interrupt. Two operational notes, both learned the hard way:

• Keep yt-dlp current. A stale build fails with "Sign in to confirm you're not a bot" on every request, which looks exactly like an IP ban but is not. brew upgrade yt-dlp (or pip install -U yt-dlp) fixes it.
• Do not raise --workers much. YouTube throttles bulk caption fetches. The --give-up-after guard aborts on sustained throttling rather than burning through the queue against a block.

Sessions whose recording genuinely has no published captions are reported as no-captions rather than as errors.

Transcribing without captions

Where no captions exist — the audio era (IETF 90-94) and the caption-less 2016-2017 videos (IETF 95-98) — the audio is transcribed on-device with Apple's Speech framework via the mi CLI (macOS 26+). It runs on the Neural Engine at roughly 90x real time, needs no model download or API key, and reads compressed audio directly. Output carries vendor: apple, product: speechanalyzer; it has segment-level timing but no per-word timestamps or confidence, so those fields are omitted rather than fabricated.

For sessions with a YouTube recording (extract audio, then transcribe):

python scripts/transcribe_mi.py --meetings 95-98

For the audio era, whose ietf.org/audio/*.mp3 recordings moved behind IETF SSO in 2026 (an anonymous request gets a Cloudflare 403 and an auth.ietf.org redirect), fetching needs a browser-grade TLS fingerprint and an authenticated session. scripts/transcribe_audio_authed.py reuses the operator's logged-in Chrome session — it never handles a password:

# 1. Sign in at datatracker.ietf.org in Chrome, then open one audio URL
#    (e.g. https://www.ietf.org/audio/ietf92/<file>.mp3) to complete the
#    get.ietf.org OIDC flow until the file plays.
# 2. Then:
pip install curl_cffi browser_cookie3   # into the ietf2vcon venv
python scripts/transcribe_audio_authed.py --meetings 90-93

It re-reads Chrome cookies on every download to ride through Cloudflare clearance rotation, and stops cleanly (resumably) if the browser session expires. Both scripts are idempotent — a session that already has a transcript is skipped — and both retry transient YouTube JS-challenge failures. Some audio-era files are Ogg Vorbis mislabelled as .mp3, which Apple's decoder rejects; the authed script transcodes those with ffmpeg and retries.

Speechmatics Transcription

This repository includes tools to re-transcribe IETF meeting audio using Speechmatics for higher-quality transcriptions with speaker diarization.

No Speechmatics transcripts are currently in the dataset. Every published transcript is either a YouTube auto-caption or, for IETF 125, local Whisper output. This section documents a tool that is available, not data you will find in the files.

Prerequisites

1. Speechmatics API Key: Sign up at speechmatics.com and obtain an API key.

2. FFmpeg: Required for audio processing.

   • macOS: brew install ffmpeg
   • Ubuntu/Debian: apt install ffmpeg
   • Windows: Download from ffmpeg.org

3. Python dependencies:

   pip install -r scripts/requirements.txt

Usage

Set your API key as an environment variable:

export SPEECHMATICS_API_KEY="your-api-key-here"

Transcribe a single vCon file:

python scripts/transcribe.py ietf121/ietf121_quic_33502.vcon.json

Transcribe all sessions from a specific meeting:

python scripts/transcribe.py --meeting 121

Transcribe a specific working group:

python scripts/transcribe.py --meeting 121 --group quic

Transcribe all vCons missing Speechmatics transcription:

python scripts/transcribe.py --all-pending

Preview which files would be transcribed:

python scripts/transcribe.py --all-pending --dry-run

Transcription Output

The script:

1. Downloads audio from the YouTube recording linked in each vCon
2. Submits the audio to Speechmatics for transcription with speaker diarization
3. Converts the result to WTF (World Transcription Format)
4. Updates the vCon file with the new transcription in the analysis array

The Speechmatics transcription is stored alongside any existing YouTube transcription, with "vendor": "speechmatics" to distinguish it.

WTF Format Features

The Speechmatics transcription includes:

• Word-level timestamps: Precise timing for each word
• Speaker diarization: Identification of different speakers
• Confidence scores: Per-word and per-segment confidence metrics
• Segments: Logical groupings of speech (sentences/phrases)
• Quality metrics: Overall transcription quality assessment

Local Whisper Transcription

This repository also includes a tool for transcribing IETF meetings using OpenAI Whisper locally via faster-whisper. No API key or cloud service is required.

Prerequisites

1. FFmpeg: Required for audio processing.

   • macOS: brew install ffmpeg
   • Ubuntu/Debian: apt install ffmpeg

2. Python dependencies:

   pip install -r scripts/requirements.txt

Usage

Transcribe a single vCon file:

python scripts/whisper_transcribe.py ietf125/ietf125_quic_XXXXX.vcon.json

Transcribe all sessions from a specific meeting:

python scripts/whisper_transcribe.py --meeting 125

Transcribe a specific working group:

python scripts/whisper_transcribe.py --meeting 125 --group quic

Use a faster (smaller) model:

python scripts/whisper_transcribe.py --meeting 125 --model medium

Transcribe all vCons missing Whisper transcription:

python scripts/whisper_transcribe.py --all-pending

Preview which files would be transcribed:

python scripts/whisper_transcribe.py --meeting 125 --dry-run

Model Selection

Model	Size	Speed	Quality
tiny	39M	Fastest	Low
base	74M	Fast	Fair
small	244M	Moderate	Good
medium	769M	Moderate	Better
large-v3	1.5G	Slow	Best (default)

Transcription Output

The script:

1. Downloads audio from the YouTube recording linked in each vCon
2. Transcribes locally using faster-whisper with word-level timestamps
3. Converts the result to WTF (World Transcription Format)
4. Updates the vCon file with the new transcription in the analysis array

The Whisper transcription is stored with "vendor": "whisper" to distinguish it from YouTube auto-captions and Speechmatics transcriptions. It includes real word-level timestamps and per-segment confidence scores.

Related Specifications

• draft-ietf-vcon-vcon-core - vCon container format (supersedes the expired draft-ietf-vcon-vcon-container)
• draft-howe-vcon-wtf-extension - World Transcription Format
• draft-howe-vcon-lawful-basis - Lawful basis extension

Contributing

Contributions are welcome! Please open an issue or pull request if you:

• Find errors in the vCon data
• Want to add vCons for additional meetings
• Have suggestions for improvements

License

This data is made available under the BSD-3-Clause License. See LICENSE for details.

The underlying IETF meeting content is subject to the IETF Trust Legal Provisions.

Acknowledgments

• IETF for making meeting recordings and materials publicly available
• The vCon working group for developing the conversation container standard
• YouTube for hosting IETF meeting recordings with auto-generated captions