Private AI for podcast show notes without sending recordings to a third party
Self-hosted speech-to-text plus an LLM can turn a recorded episode into a transcript, timestamped chapters, a summary, quotable lines and a feed description, without uploading unreleased audio to a transcription vendor. By hand, that's a couple of hours a week nobody has spare. The raw recording is also more sensitive than it looks: a guest who hasn't approved a segment, an announcement floated off the cuff, a remark meant to stay off the record until publication.
What's in an unreleased recording
A rough cut carries more than the finished episode does. Guests say things in the room they'd want cut before publishing, hosts float announcements that aren't confirmed yet, and a sensitive topic might get discussed candidly with the understanding that only the edited version goes out. Most of it isn't secret in a legal sense, but it is pre-release material a show can leak if the raw audio or transcript sits in a vendor's storage under retention terms nobody on the team read closely.
What running the model yourself changes
A self-hosted speech-to-text model paired with an LLM for summarization runs the whole pipeline (audio in; transcript, chapters, summary and show notes out) on models the show runs itself. No recording goes to a transcription vendor and back. On a rented Spark the machine is GPUwerk's hardware, assigned only to you and operated under its published DPA.
Output quality is a separate question. A good open-weight speech model handles typical podcast audio (two or three speakers, moderate background noise) reasonably well, and the LLM step turns the transcript into chapters and notes in the show's usual format.
This is the same self-hosted speech pipeline covered more generally in self-hosted speech and voice AI; show notes are one specific application of it.
A concrete example
A weekly interview show records a 50-minute episode. A self-hosted transcription model produces a full transcript with speaker labels, then an LLM pass generates five timestamped chapter markers, a 150-word episode summary for the feed description, and a shortlist of quotable lines the host can pull for social promotion. The editor reviews the transcript against the actual audio for anything the model misheard, particularly names and technical terms, confirms the summary doesn't reveal anything the guest asked to keep out of the public description, and publishes. What used to take an editor most of an afternoon now takes a review pass over model output.
Where this is not a drop-in replacement
Automated transcription still misses names, jargon, and cross-talk, and a model summarizing an episode has no sense of what a guest considers embarrassing; that judgment stays with the host or editor who was in the conversation. Timestamped chapters generated automatically are a starting point, not a final cut; someone who listened to the episode should confirm the chapter boundaries land where a listener would expect them. Treat the pipeline as a fast first draft of the notes, not a replacement for an editor who knows the show and the guest.
One Spark covers a weekly show
Speech-to-text and summarization fit side by side on a single Spark within 128 GB of unified memory, and a $0.79/hour on-demand instance covers a weekly show's processing without needing a reserved instance running between episodes. See the setup guide for getting a Spark running from a fresh instance.