---
license: other
license_name: npr-fresh-air-transcripts
license_link: https://github.com/whit3rabbit/freshair-analysis/blob/main/LICENSE
tags:
- npr
- podcast
- interview
- emotion-analysis
- nlp
pretty_name: Fresh Air Interview Corpus (2006-2026)
---

# Fresh Air interview corpus

NPR *Fresh Air* interview transcripts (2006-2026, ~1,070 episodes) plus derived
training/verification data, used by
[whit3rabbit/freshair-analysis](https://github.com/whit3rabbit/freshair-analysis) - an analysis
pipeline and a Claude Code "interviewer coach" skill trained on Terry Gross's measured
interviewing behavior.

## License note

The pipeline *code* in the GitHub repo is 0BSD. This dataset is NPR-sourced transcript content
and derived annotations, not covered by that license - it's shared here for research and
reproducibility of the linked analysis, not as a redistribution license for the underlying NPR
material.

## Layout

```
freshair_data/
  article_urls.json                    URL index for each episode
  full_transcripts.jsonl                raw scraped transcripts
  dialogue.jsonl                        turn-segmented dialogue
  dialogue_clean.jsonl                  id-repaired version of dialogue.jsonl (see repo CLAUDE.md)
  gross_questions.jsonl                 Terry Gross's turns, isolated
  freshair_training_dialogue_acts.jsonl dialogue-act-labeled training data
  freshair_training_verified.jsonl      MiniMax-verified emotion labels

analysis_data/
  freshair_training_dialogue_acts.jsonl  (identical copy - analysis/ pipeline reads from here)
  freshair_training_verified.jsonl       (identical copy - analysis/ pipeline reads from here)
  freshair_training_features.jsonl       final ML-ready feature table (852/107/107 episode split)
```

See the GitHub repo's `analysis/data/README.md` for the authoritative per-file schema, and
`CLAUDE.md` for how this corpus was built and recovered.