- youtube_transcript_api to fetch the transcript
- textblob to compute sentiment
- openai api to build a a summary (needs API KEY & purchase tokens)
- Summarize with local LLM i/o burning OPENAI_API credits (command-line switch)
- Rather than relying on YT transcripts of dubious quality, download YT audio w. yt_dlp and process locally with whisper
- Use spacy & nltk to improve transcript quality. Extend further by:
- Integrating a Levenshtein distance on phoneme sequences
- Prioritizing corrections based on contextual language models (e.g., transformers)
- Filtering using part-of-speech or NER if needed
- Use PyAnnote so that it recognizes speakers (a.k.a. speaker diarization) so as to be able to summarize separately their opinions / rationale from a conversation