loading
loading
Exa Opens 3.2M Podcast Transcripts to Developers Through One Search API Exa added 3.2 million podcast episodes with full transcripts to its search API, letting agents query 2.2 million hours of audio for free. - Exa added 3.2 million podcast episodes with full transcripts to its search index. - Corpus covers 119,000 shows, 23.7 billion words, roughly 2.2 million hours of audio. - Returns newest episode in top 5 results 98.4% of the time across 250 shows benchmarked. - Answers 85.0% of 667 episode specific questions correctly, beating six competing APIs. - Transcripts include speaker labels, publish dates, hosts, guests, and show metadata. - Available now across all Exa endpoints at no additional charge via the dashboard. Exa has added full podcast transcripts to its web search API, giving developers a way to search spoken comments through the same endpoint used for web results. According to the company’s launch announcement, the index covers 3.2 million episodes from 119,000 shows, totaling about 23.7 billion words and 2.2 million hours of audio. The transcript layer exposes material that conventional search often misses. A CEO’s pricing comments in a two-hour interview, for example, can now appear in an agent’s search results even when the episode title and show notes never mention pricing. Exa says it adds new episodes soon after publication. Inside the podcast corpus The index spans technology, business, finance, news, science, culture, and sports. Queries can match the words spoken during an episode alongside its title, description, and other metadata. - Show and episode titles, publisher, and description - Hosts, guests, publication date, and duration - Full transcripts with passages assigned to individual speakers Speaker labels make the transcripts more useful for automated research. An application can isolate a guest’s statements instead of treating every mention of that person’s name as relevant. Exa’s benchmark edge Exa compared its service with six other podcast search APIs using two tests: whether search surfaced the latest episode of an active show, and whether results supported the correct answer to an episode-specific question. Because Exa ran and published the benchmark, developers should treat it as a vendor result until an independent evaluation reproduces it. The figures still provide useful expectations: freshness was consistently high, while about 15% of episode-specific questions produced an incorrect result. Three practical workflows Searchable transcripts can support applications that need direct quotations, recent commentary, or statements attributed to a specific person and date. - Catch-up assistants: Summarize a recent episode, answer questions about it, or retrieve passages from a named speaker. - Market and investment research: Track comments about demand, pricing, hiring, or competition between formal company disclosures. - Sales and recruiting preparation: Find a prospect’s or candidate’s recent appearances and extract relevant views before a meeting. One endpoint, no new SDK Existing integrations can continue using Exa’s /search route and normal contents payload. A query such as podcast episodes where founders discuss pricing mistakes can return podcast results and transcripts as structured data for downstream parsing, retrieval, or summarization. Exa says the podcast index is available to all API users without a separate podcast surcharge, and queries can also be tested in its dashboard playground. The release follows the company’s Exa 3.0 search model and extends the same search infrastructure to another content type. Production caveats The reported 85% correctness rate leaves enough retrieval errors to require verification for financial, legal, medical, or other high-stakes uses. Generated answers also need citations to the underlying transcript passages so users and systems can inspect the supporting language. Exa’s announcement does not quantify transcription accuracy, language coverage, indexing latency, or speaker-assignment errors. Noisy recordings, overlapping dialogue, panel discussions, and non-English episodes can all reduce transcript quality or attach words to the wrong person. Newly published audio may also remain unavailable until Exa processes and indexes it. For research agents, sales tools, and personal assistants, the release opens a large body of spoken material to the same retrieval workflow already used for web pages. Production systems should pair that access with transcript citations, confidence checks, and source verification.
document fetched 2026-10-29 03:02