Podcast Transcriber & Monitor — Backfill + New Episode Alerts

Turn any podcast RSS feed into searchable text — in bulk, or on autopilot.

Give it an RSS feed and it transcribes the back catalog (oldest episodes included), or keeps watch and transcribes every new episode as it drops. Two speech engines work together: a frontier ASR model writes the transcript, and a second engine adds speaker labels ("who said what") on the Speaker tier.

What you get per episode

• Full transcript with timestamps, plus ready-to-use SRT and VTT subtitle files
• Speaker labels on the Speaker tier, with the host's split labels merged automatically and an optional name mapping (give it ["Host", "Guest"] and get readable dialogue)
• Show notes context: episode title, publish date, duration, description and link travel with every transcript
• A run summary in the key-value store (OUTPUT): episodes processed, billed minutes, and per-engine usage for cost reconciliation

Two modes

Mode	What it does
backfill	Transcribe existing episodes from a feed. Defaults to the latest 10 episodes so a 400-episode archive never surprises your wallet. Pick by latest N, date range, or all — with hard caps you control.
monitor	Remembers what it has already transcribed (stored in the key-value store, keyed per feed). Schedule it hourly or daily; only genuinely new episodes are transcribed and billed. An optional webhook pings your endpoint when new transcripts land.

Feeds are fetched politely with conditional requests (ETag / Last-Modified): an unchanged feed costs one cheap feed-check and zero transcription.

Pricing (pay per event)

Event	Price
Standard transcript (per audio minute, rounded up)	$0.010
Speaker-labeled transcript (per audio minute, rounded up)	$0.018
Feed check (monitor mode, per feed per run)	$0.002
Actor start	$0.0005

• Your first episode is free — the first standard-tier episode of a run (up to 60 minutes) is on us.
• Failures are never billed. Unreachable audio, unsupported content, silence — if there's no transcript, there's no transcription charge.
• Duplicates in monitor mode are never billed.

Why two engines?

Because one model can't do both jobs best. The primary engine produces the cleanest transcript text (it's the one that hears "Stairway to Heaven", not "steroid to heaven"), while the second engine is the one that can tell speakers apart. On the Speaker tier both run and their outputs are aligned by timestamp, so you get the best text with speaker labels. If a podcast CDN blocks one engine, the other takes over automatically — you still get your transcript.

Input

• feedUrls — one or more RSS feed URLs (Apple Podcasts show links are resolved automatically), or audioUrls — direct audio file URLs
• mode — backfill (default) or monitor
• episodeSelection — latestN (default 10), afterDate, beforeDate, or all
• tier — standard (transcript only) or speaker (transcript + speaker labels)
• language — auto (default), en, zh, and more
• speakerNames — optional names mapped onto speakers in order of appearance
• Advanced: freeFirstEpisode, webhookUrl, maxEpisodesPerRun (default 200), maxEpisodeMinutes (default 240)

Output

One dataset row per episode: status, showTitle, episodeTitle, publishedAt, audioUrl, durationSec, language, transcript, segments (start/end/speaker/text), speakers, srt, vtt, wordCount, billedMinutes, enginesUsed — plus the OUTPUT summary record in the key-value store.

Notes & limits

• Episodes longer than maxEpisodeMinutes (default 4 hours) are skipped, never billed.
• Please only transcribe content you have the rights to process.
• Very old episodes behind multi-hop redirect chains are resolved to direct links automatically; the rare file both engines cannot fetch is reported as failed (with the reason), never billed, and never filled in with made-up text.

---

Built for podcasters, researchers, and content teams who need their audio searchable.