PPT Audio To Video

Use this skill when the source video has narration audio but no usable slide visuals, and the final deliverable should be a slide-based lecture video.

Resolve bundled scripts relative to this skill directory. If the runtime has already opened this SKILL.md, prefer paths like scripts/extract_slide_outline.py and scripts/render_from_timing_csv.py instead of machine-specific absolute paths.

Core workflow

1. Inventory inputs.

   • Confirm which of these exist: audio-only mp4/m4a/mp3/wav, ppt/pptx, pdf, and any pre-rendered slide images.
   • Prefer an existing pdf or image directory for rendering. Treat pptx as the source of slide text and as a fallback for export.

2. Prepare tools.

   • Required for deterministic steps: ffmpeg, ffprobe, pdftoppm.
   • Required for transcription: whisper-cli from whisper-cpp plus a multilingual model such as ggml-small.bin.
   • If only pptx exists and no pdf/images exist, prefer Keynote or PowerPoint export on macOS. Use soffice only as fallback because profile or rendering issues are common.

3. Produce slide images.

   • If pdf exists, render it to images:

     pdftoppm -png -r 200 "$PDF" "$OUTDIR/slide"
     

   • If only pptx exists, export to pdf or slide images with Keynote or PowerPoint, then continue from pdf.
   • Keep slide filenames ordered and stable, such as slide-01.png, slide-02.png, ...

4. Extract slide text.

   • Run:

     python3 scripts/extract_slide_outline.py \
       --pptx "$PPTX" \
       --out "$WORKDIR/slide_outline.csv"
     

   • Use the output to identify slide titles, distinctive keywords, and section changes.

5. Extract clean audio for ASR.

   • For audio-only mp4, extract mono wav:

     ffmpeg -y -i "$AUDIO_MP4" -ar 16000 -ac 1 -c:a pcm_s16le "$WORKDIR/audio.wav"
     

   • If the source is already wav/mp3/m4a, convert to the same mono wav form if needed.

6. Transcribe with whisper-cli.

   • Example:

     whisper-cli -ng \
       -m "$MODEL" \
       -f "$WORKDIR/audio.wav" \
       -l zh \
       -ocsv -osrt -of "$WORKDIR/transcript"
     

   • Prefer transcript.csv for downstream parsing. transcript.srt is useful for manual review.
   • If GPU allocation fails on macOS, retry with -ng to force CPU mode.

7. Build slide_timings.csv.

   • Do not average slide durations unless the user explicitly asks for it.
   • Read the transcript and slide outline together, then create a monotonic timing plan by topic changes, section boundaries, and unique keywords.
   • Use this schema:

     slide,start_sec,end_sec,duration_sec,reason
     1,0.000,15.000,15.000,opening title and agenda
     2,15.000,100.000,85.000,architecture overview starts here
     

   • Keep slide numbers sequential and ensure duration_sec = end_sec - start_sec.
   • Validate that the last end_sec matches the audio duration or is within a small tolerance.

8. Render the final video.

   • Run:

     python3 scripts/render_from_timing_csv.py \
       --images "$SLIDE_IMAGES_DIR" \
       --timings "$WORKDIR/slide_timings.csv" \
       --audio "$WORKDIR/audio.wav" \
       --output "$OUT_VIDEO"
     

   • The script generates an ffconcat file, validates timing continuity, and calls ffmpeg to encode the final mp4.

9. Verify and iterate.

   • Check output duration with ffprobe.
   • If a slide cuts too early or too late, edit only the affected rows in slide_timings.csv and rerun the render script.
   • Keep the transcript, outline, and timing CSV as reproducible working files.

Heuristics for timing alignment

• Use section-divider slides briefly. These slides usually hold for 5-20 seconds.
• Use the first segment that clearly switches topic as the next slide start.
• Prefer exact topic transitions over title-word matching. ASR often distorts proper nouns and product names.
• Let the model infer timings, but keep the render step deterministic through slide_timings.csv.
• When confidence is low, produce a first-cut video and tell the user which slide boundaries likely need review.

Common commands

Install dependencies on macOS if missing:

brew install ffmpeg poppler whisper-cpp

Typical multilingual model download:

mkdir -p .models
curl -L 'https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-small.bin' -o .models/ggml-small.bin

Bundled scripts

• scripts/extract_slide_outline.py Extract slide text from pptx into CSV or JSON for timing analysis.
• scripts/render_from_timing_csv.py Validate a timing CSV, generate an ffconcat, and render the final video with ffmpeg.