Gettr Transcribe + Summarize (MLX Whisper)

Quick start (single command)

Run the full pipeline (Steps 1–3) with one command:

bash scripts/run_pipeline.sh "<GETTR_POST_URL>"

To explicitly set the transcription language (recommended for non-English content):

bash scripts/run_pipeline.sh --language zh "<GETTR_POST_URL>"

Common language codes: zh (Chinese), en (English), ja (Japanese), ko (Korean), es (Spanish), fr (French), de (German), ru (Russian).

This outputs:

./out/gettr-transcribe-summarize/<slug>/audio.wav
./out/gettr-transcribe-summarize/<slug>/audio.vtt

Then proceed to Step 4 (Summarize) to generate the final deliverable.

Workflow (GETTR URL → transcript → summary)

Inputs to confirm

Ask for:

GETTR post URL
Output format: bullets only or bullets + timestamped outline
Summary size: short, medium (default), or detailed
Language (optional): if the video is non-English and auto-detection fails, ask for the language code (e.g., zh for Chinese)

Notes:

This skill does not handle authentication-gated GETTR posts.
This skill does not translate; outputs stay in the video's original language.
If transcription quality is poor or mixed with English, re-run with explicit --language flag.

Prereqs (local)

mlx_whisper installed and on PATH
ffmpeg installed (recommended: brew install ffmpeg)

Step 0 — Pick an output directory

Recommended convention: ./out/gettr-transcribe-summarize/<slug>/

Extract the slug from the GETTR post URL (e.g., https://gettr.com/post/p1abc2def → slug = p1abc2def).

Directory structure:

./out/gettr-transcribe-summarize/<slug>/audio.wav
./out/gettr-transcribe-summarize/<slug>/audio.vtt
./out/gettr-transcribe-summarize/<slug>/summary.md

Step 1 — Extract the media URL and slug

Preferred: fetch the post HTML and read og:video*.

python3 scripts/extract_gettr_og_video.py "<GETTR_POST_URL>"

This prints the best candidate video URL (often an HLS .m3u8) and the post slug.

Extract the slug from the URL path (e.g., /post/p1abc2def → p1abc2def) to create the output directory.

Important: Streaming URLs require browser extraction

For streaming URLs (gettr.com/streaming/<slug>), the Python script may return a stale/invalid og:video URL that fails with HTTP 412. This is because GETTR dynamically generates signed stream URLs via JavaScript.

If the URL is a streaming link OR if download fails with HTTP 412:

Open the streaming URL in a browser and wait for the page to fully load (JavaScript must execute)

Extract the og:video meta tag content from the rendered DOM:

document.querySelector('meta[property="og:video"]').getAttribute('content')

Use that fresh URL for the download step

The browser-extracted URL will have a valid signature and work with ffmpeg.

If extraction fails, ask the user to provide the .m3u8/MP4 URL directly (common if the post is private/gated or the HTML is dynamic).

Step 2 — Download audio with ffmpeg

Extract audio-only (16kHz mono WAV) for faster and more stable transcription:

bash scripts/download_audio.sh "<M3U8_OR_MP4_URL>" "./out/gettr-transcribe-summarize/<slug>/audio.wav"

This directly extracts audio without intermediate video, reducing disk I/O and processing time.

Step 3 — Transcribe with MLX Whisper

Generate VTT output with timestamps:

mlx_whisper "./out/gettr-transcribe-summarize/<slug>/audio.wav" \
  -f vtt \
  -o "./out/gettr-transcribe-summarize/<slug>" \
  --model mlx-community/whisper-large-v3-turbo \
  --condition-on-previous-text False \
  --word-timestamps True

To explicitly set the language (recommended when auto-detection fails):

mlx_whisper "./out/gettr-transcribe-summarize/<slug>/audio.wav" \
  -f vtt \
  -o "./out/gettr-transcribe-summarize/<slug>" \
  --model mlx-community/whisper-large-v3-turbo \
  --condition-on-previous-text False \
  --word-timestamps True \
  --language zh

Flags explained:

-f vtt: VTT format provides timestamps for building the outline.
--condition-on-previous-text False: prevents hallucination errors from propagating across segments.
--word-timestamps True: more precise timing for section boundaries.
--language <code>: explicit language code (e.g., zh, en, ja, ko). Use when auto-detection fails.

Notes:

By default, language is auto-detected. For non-English content where detection fails, use --language.
Common language codes: zh (Chinese), en (English), ja (Japanese), ko (Korean), es (Spanish), fr (French), de (German), ru (Russian).
If too slow or memory-heavy, try smaller models: mlx-community/whisper-medium or mlx-community/whisper-small.
If quality is poor, try the full model: mlx-community/whisper-large-v3 (slower but more accurate).
If --word-timestamps causes issues, omit it (the pipeline script handles this automatically).

Step 4 — Summarize

Write the final deliverable to ./out/gettr-transcribe-summarize/<slug>/summary.md.

Pick a summary size (user-selectable):

Short: 5–8 bullets; (if outline) 4–6 sections
Medium (default): 8–20 bullets; (if outline) 6–15 sections
Detailed: 20–40 bullets; (if outline) 15–30 sections

Include:

Bullets (per size above)
Optional timestamped outline (per size above)

Timestamped outline format (default heading style):

[00:00 - 02:15] Section heading
- 1–3 sub-bullets

When building the outline from VTT cues:

Group adjacent cues into coherent sections.
Use the start time of the first cue and end time of the last cue in the section.

Bundled scripts

scripts/run_pipeline.sh: full pipeline wrapper (Steps 1–3 in one command)
scripts/extract_gettr_og_video.py: fetch GETTR HTML and extract og:video* URL + post slug (with retry/backoff)
scripts/download_audio.sh: download/extract audio from HLS or MP4 URL to 16kHz mono WAV

Error handling

Non-video posts: The extraction script detects image/text posts and provides a helpful error message.
Network errors: Automatic retry with exponential backoff (up to 3 attempts).
No audio track: The download script validates output and reports if the source has no audio.
HTTP 412 errors: The extracted og:video URL has an expired/invalid signature. Use browser extraction to get a fresh URL (see Step 1 and references/troubleshooting.md).

Troubleshooting

See references/troubleshooting.md for detailed solutions to common issues including:

HTTP 412 errors (stale signed URLs)
Extraction failures
Download errors
Transcription quality issues