space-tape

Open-source X Spaces transcription.

The page is the git.

Give it a Spaces link (or the tweet that posted the replay). Get a timed transcript with speakers, plus the audio file. Keep the original X link next to the result.

Whisper is the ear. Hydra is the speaker list.

Demo

The Space @elonmusk posted on 24 May 2023 — DeSantis announcement, servers straining, ~667,000 trying to get in.

https://x.com/i/spaces/1eaJbrAlZjjJX

Play the tape on the Space Tape site (this git as a page). Cues live in examples/1eaJbrAlZjjJX. Audio is the real replay. Hydra lives in the raw mpegts — remux strips it.

pip install git+https://github.com/hovtgc/x-spaces-transcription.git
space-tape transcribe "https://x.com/elonmusk/status/1661498079386931206"

Package name: space-tape. You get cues.json, transcript.md, and audio.mp3. No browser. No X API key. No virtual computer.

Repo: hovtgc/x-spaces-transcription

Do you need a virtual computer?

No — not for recorded Spaces.

A replay is just a file. Three things come out of it:

Signal	Where it lives	Tool
Words	mixed audio	Hugging Face Whisper (`transformers` ASR pipeline)
Who is unmuted	ID3 tags muxed into the HLS/m4a (`HydraParticipants`, `HydraAudioLevel`, NTP)	`space_tape/hydra.py`
Identity (handle, pfp, X user id)	those same Hydra tags	no API call required

A browser agent watching the unmute UI is a fallback, not the pipeline.

If you have the replay as downloaded (.m4a / .ts from yt-dlp), skip the computer. An agent only needs a shell (yt-dlp, ffmpeg, Python).

Pipeline

Spaces or tweet URL
        │
        ▼
 yt-dlp ──► replay.m4a / replay.ts   (keep ID3)
        │
        ├─────────────────────────────┐
        ▼                             ▼
 parse Hydra ID3              ffmpeg → 16 kHz mono wav
 (who's unmuted, ~1 Hz)               │
        │                             ▼
        │                     Hugging Face Whisper
        │                     (segments + optional words)
        ▼                             ▼
        └──────────── merge ──────────┘
                      │
                      ▼
         cues.json + transcript.md + audio.mp3
         (and the original X URL, untouched)

See AGENTS.md for the agent playbook.

Quick start

Accept any of:

https://x.com/i/spaces/1pKdRDlVbRrJW
https://x.com/notpierce69/status/2100117423017906497
https://twitter.com/i/spaces/...

# KEEP the mpegts/m4a. Do not remux yet — remux strips Hydra tags.
space-tape transcribe "https://x.com/i/spaces/1pKdRDlVbRrJW" -o ./out
space-tape transcribe "https://x.com/user/status/123" --model openai/whisper-small.en

Python:

from space_tape.download import download, parse_url
from space_tape.hydra import hydra_series, regions
from space_tape.merge import attach_speakers
from space_tape.transcribe import transcribe
from space_tape.render import wav_for_asr, write_outputs

System: ffmpeg. Whisper weights: Apache 2.0.

English Spaces: openai/whisper-small.en or medium.en. Mixed language: openai/whisper-large-v3.

Hydra

X's audio stack is called Hydra. About once a second it writes ID3v2 frames into the stream:

Frame	Description
`TXXX JSONMetadata`	`{ "HydraVersion": 4, "ntp": … }` — clock
`TXXX HydraParticipants`	JSON array of people on stage (not listeners)
`TXXX HydraAudioLevel`	JSON int array, one level per on-stage slot

Index mapping that worked in practice:

HydraAudioLevel[0] = host
HydraAudioLevel[i+1] = HydraParticipants[i]
level 0 = muted / silence
level ≳ 8 = unmuted and making sound

NTP is seconds. Subtract the first NTP to get media time t.

Critical: parse the raw yt-dlp file. ffmpeg -c copy to .aac or transcoding to wav drops every Hydra tag. Parse first, then convert a copy for Whisper.

Host is not in HydraParticipants. Bake the host handle from the tweet author or --host.

Drop a raw replay on the Player tab (or space-tape hydra replay.m4a). Nothing is uploaded.

Output

cues.json — this is the product.

[
  {
    "start": 5.004,
    "end": 7.003,
    "speaker": "notpierce69",
    "text": "Hey DJ, I don't think anyone else is gonna show up."
  }
]

transcript.md keeps the X URL. Playback on a site is a convenience. X is the source.

Suggested player: one <audio> + the cue list. Highlight the cue whose [start, end] contains audio.currentTime.

Gotchas (paid for in blood)

Never transcode before parsing ID3. The .aac/.wav will transcribe and have zero speakers.
Hydra ticks ~1 Hz, Whisper segments are 2–8 s. Midpoint assignment is right 95%+ of the time. Don't overfit.
Host is slot 0 of HydraAudioLevel.
HydraParticipants is on-stage only. Listeners never appear.
Empty [] participants + [0] levels = nobody talking, or host muted.
yt-dlp + X will break. Pin yt-dlp, document the last working version.
Replays expire. Download is the archive.
Whisper hallucinates on long silence. Skip Hydra-silence regions if you need to.
Display names change. Bake the handle.

Ethics / ToS

Public recorded Spaces only.
Keep the link to X. You are making a transcript, not a mirror.
Don't ship other people's Spaces as if you hosted them.
Whisper weights: Apache 2.0 (openai/whisper). Hydra tags are in a file the user already received from X.

MIT. Use it, fork it, host the git as a page.