Podcast Chapter Generator using Whisper

Podcast episodes routinely run one to three hours, yet most players offer nothing better than a scrub bar — listeners cannot jump to the interview segment or skip the sponsor read. Creators know chapters fix this, but hand-writing timestamps for every episode is tedious enough that most skip it. Speech recognition changed the economics: Whisper produces timestamped transcripts good enough to work from, and topic shifts in the transcript reveal natural chapter boundaries. This project wires that into a complete pipeline — transcribe, segment, title, review on a waveform, export — so a creator uploads an episode and gets ready-to-paste chapters. The listing is careful about what is automated and what is not: transcription quality depends on the audio, and chapter boundaries are heuristic suggestions the user reviews, not perfect cuts.

1. The user uploads an MP3/WAV episode; FFmpeg normalizes it to the 16 kHz mono format Whisper expects.
2. faster-whisper runs the selected pre-trained Whisper model and returns segments with start/end timestamps.
3. A segmentation stage combines long silences with topic-shift signals in the transcript to propose chapter boundaries near the target length.
4. Each proposed chapter receives an extractive title from its most distinctive phrases.
5. The web UI renders the waveform with chapter regions; the user drags boundaries, splits, merges and retitles as needed.
6. Export writes YouTube-format chapter text, MP3 chapter tags and JSON from the reviewed chapter list.

Tech stack:

• Python 3.10, faster-whisper (CTranslate2 runtime)
• OpenAI Whisper pre-trained weights (open-source; downloaded once)
• FFmpeg (audio normalization and chapter tagging)
• Flask backend, HTML/CSS/JS waveform editor frontend

• [Whisper transcription] Pre-trained Whisper models (tiny through large, selectable) transcribe any episode with segment-level timestamps; no model training is required.
• [Automatic chapter segmentation] Long silences and topic shifts in the transcript are combined to propose chapter boundaries, with a configurable target chapter length.
• [Extractive chapter titles] Each proposed chapter gets a title drawn from its most distinctive phrases — transparent and inspectable, not a black-box generator.
• [Interactive waveform editor] The episode renders as a chapter-colored waveform; boundaries can be dragged, chapters split, merged and retitled before export.
• [Multi-format export] One click produces YouTube description chapters, MP3 ID3 chapter tags and a JSON chapter file for other players.
• [Batch queue] Drop in several episodes and process them sequentially, with per-episode status and logs.
• [Configurable pipeline] Model size, language hint, silence thresholds and target chapter length are all adjustable from the UI.

• Complete source code (transcription pipeline, segmentation, web UI, exporters)
• Setup guide (Python environment, Whisper weight download, FFmpeg, running the UI)
• Project report PDF (background, Whisper overview, segmentation method, evaluation procedure)
• PPT presentation for final review
• Viva Q&A preparation document (speech recognition, transformers, timestamp alignment, segmentation heuristics)
• Demo guidance with a sample episode workflow

• Transcription errors rise with heavy accents, crosstalk, background music and poor microphones — the UI is built for reviewing, not blind trust.
• Chapter boundaries are heuristic suggestions; topic shifts in casual conversation do not always match, so human review matters.
• Large Whisper models need significant RAM/VRAM and time on CPU; the tiny/base models are provided as fast fallbacks.
• Very long episodes (3+ hours) need patience and disk space for intermediate files.
• The pipeline does not separate overlapping speakers; it chapters topics, not diarized speakers.