Transcribe Note

Ocupación

Categorías

Medios

find ~/Downloads -maxdepth 1 -type f \( -name '*.m4a' -o -name '*.mp3' -o -name '*.wav' -o -name '*.ogg' -o -name '*.flac' -o -name '*.aac' -o -name '*.opus' -o -name '*.webm' -o -name '*.mp4' -o -name '*.wma' \) -print0 | xargs -0 ls -t 2>/dev/null | head -1

If no audio file is found, tell the user and stop.

Print the filename and file size so the user can confirm the right file was found.

2. Transcribe with whisperx

Run whisperx with the large-v3 model and int8 compute type (required on Apple Silicon CPU), outputting plain text to a temporary directory:

TMPDIR=$(mktemp -d)
whisperx --model large-v3 --compute_type int8 --output_dir "$TMPDIR" --output_format txt [--language LANG] "AUDIO_FILE"

Read the resulting .txt file from $TMPDIR. Clean up the temporary directory afterwards.

If whisperx fails, do not fall back to whisper (it runs on CPU without ctranslate2 and is extremely slow). Instead, diagnose the failure: spawn a subagent to investigate the error, identify the root cause (e.g. compute type incompatibility, broken ffmpeg linkage, missing libraries), and fix it. Then retry whisperx. Only proceed to step 3 once whisperx succeeds.

3. Ask the user for a title

Use AskUserQuestion to ask the user what the note should be titled. Show them the first ~200 words of the transcript so they have context.

4. Create the note

Using the full transcript as source material, write a well-organized org-mode note. This is not a verbatim copy of the transcript. Instead:

• Restructure the content into logical sections with org headings (** level, since the top-level * heading is the title).
• Clean up verbal artifacts: filler words, false starts, repetitions, digressions.
• Preserve all substantive content, ideas, arguments, and nuances. Do not omit points just because they were expressed informally.
• Improve the prose so it reads like a well-written document, while keeping the author's voice and intent.
• Add structure where the spoken content was stream-of-consciousness: group related points, add section headings, use lists where appropriate.

The note should follow the org-note format:

1. Slugify the title. Run emacsclient -e '(simple-extras-slugify "TITLE")' to get the slug. Use the result (minus surrounding quotes) as the filename, appending .org.

2. Write the file to /Users/pablostafforini/My Drive/notes/SLUG.org with this structure:

   #+title: TITLE
   
   * TITLE
   RESTRUCTURED CONTENT HERE
   

3. Generate an org ID. Run:

   emacsclient -e '(with-current-buffer (find-file-noselect "FILEPATH") (goto-char (point-min)) (org-next-visible-heading 1) (org-id-get-create))'
   

5. Confirm

Tell the user the note has been saved, showing the file path and a brief summary of the sections created.