From a live X Space to a full transcript in two languages

The pipeline

Five steps, all on one ordinary Windows PC with WSL:

1. Record. Audacity, with the Windows WASAPI host and the speaker output as the input device ("loopback"). It records whatever the PC plays, so the Space just runs in the browser. Export as MP3.
2. Transcribe. faster-whisper with the small.en model, on the CPU. It writes an .srt file with timestamps and a plain text file.
3. Clean. Claude reads the raw transcript and writes a clean English version: speaker names, paragraphs, misheard words fixed, filler removed.
4. Translate. A full Turkish translation of the clean English text. Not a summary.
5. Publish. Both versions as X Articles, plus PDF and DOCX copies for anyone who wants a file.

Why record live instead of downloading the replay? A replay is not always kept, and it may arrive hours later. A local recording is under my control from the first second.

Cleaning is the real work

The raw Whisper output was readable but not publishable. The typical problems:

• Names. Whisper doesn't know who is talking and often mishears names. The host's name needed fixing throughout, and every paragraph needed a speaker label.
• Numbers. In crypto, a wrong number is worse than a missing one. Every figure (for example the share of tokens in the team multisig) was checked against the audio, not guessed from context.
• Tone. Spoken English is looser than written English. A few rough words were softened in both versions without changing the meaning.
• Honesty. Both texts start with a note: transcribed from an audio recording, so a few words may be wrong. Readers deserve to know what kind of text they are reading.

The clean English text came to about 8,200 words. The Turkish translation is shorter in word count, as Turkish usually is.

Where it went wrong

1. My first idea for X was wrong. I wanted the Turkish text as a thread of images. The pages came out fine (36 of them), but uploading them through the Typefully API hit its limit of 20 requests per hour. One article is better than 36 images anyway: it is searchable, it can be copied, and it works with screen readers. So we switched to an X Article.

2. X has no PDF attachments. A post can carry up to 4 images, one GIF or one video. A PDF has to live somewhere else and be linked. The X Article made this moot.

3. The browser translated my English article. I pasted the English article from a page opened in Edge. Edge had auto-translated that page into Turkish, so the English article went live with a Turkish title. Claude caught it by reading the published post back, not by trusting that the paste had worked. The fix took one edit. The lesson: turn off automatic translation on any page you copy text from.

What I'd keep

• Record locally, even if a replay will exist.
• Use Whisper for the first draft only. The clean version needs a careful pass, especially for names and numbers.
• Translate the full text, not a summary. People who read the translation should get the same AMA as people who listened.
• Check what was actually published, not what you think you published.

The articles are on my X profile: the English transcript and the Turkish translation.