Turn Google Drive scripts into narrated audio, line by line

Build me an audio production desk for the team that narrates our content with Hume Octave. It is a two pane workspace: a list of narration scripts on the left, and the selected script opened as an editable segment table on the right. Producers live in this screen, so it has to be fast to scan and safe to redo one thing at a time.

Left pane, the script library. A handler calls Google Drive List Files scoped to a configurable folder id, filtered to Google Docs and plain text files that are not trashed, and shows each script with its name, last modified date, and a status badge. Status is one of draft, generated, or approved, stored by the app against the Drive file id so the team always knows what is still outstanding. Let me filter the list by status and search by name. Selecting a script loads its text: use Export Google Workspace File for anything with the Google Docs mime type, and Download File Content for plain text files.

Right pane, the segment table. Each row is one narration segment and carries the segment text (editable), an assigned voice picked from a dropdown populated by Hume List Voices (include both the shared Voice Library and our own custom voices), an optional acting instruction describing the delivery, a per row Generate button, an inline audio player for the latest take, and a small state marker showing whether that row has been generated. The Generate button calls Hume Text-to-Speech (JSON) for that one segment only, passing the segment text, the chosen voice, and the acting instruction as the description. Use the JSON variant rather than the file variant on purpose, because the JSON response carries base64 audio the app can play inline plus the metadata we need to keep. Also give me a Generate all button that runs ungenerated rows as a paced batch with a visible progress bar.

Persist each segment's generation id from the text to speech response alongside its audio. This is load bearing: Hume's Save Custom Voice only accepts a generation id returned by an earlier text to speech call, so if the app throws the id away, promoting a take to a reusable voice becomes impossible. Keep the id on the row even after regeneration, replacing it with the id of the newest take.

Add a Prep this script button that kicks off a background agent. The agent reads the selected Drive document, splits it into natural narration segments, works out who is speaking in each one, assigns a voice per speaker or character from the voices returned by Hume List Voices, and drafts a short acting instruction for each segment describing tone and pacing. It writes all of that back into the segment table as a proposal. It must not generate any audio. A human reviews the split, swaps voices, and edits the instructions before pressing Generate on anything.

Approving a script runs the publish step. Create a folder for the project with Google Drive Create Folder under a configurable parent, upload the generated segment audio into it with Upload File (Multipart) named by segment order so the files sort correctly, then post a message to a configurable Slack channel with Slack Bot Send a Message containing the script name and the Drive folder link. Note that Upload File (Multipart) covers files up to 5MB, so upload per segment rather than as one giant file. Slack has no file upload in this catalog, so the audio is shared as a Drive link rather than an attachment. Approving flips the script status to approved.

Give me a voice library panel too. From any generated row, a Save as brand voice action calls Hume Save Custom Voice using that row's retained generation id and a name I type, and the new voice immediately becomes selectable in the voice dropdown. The panel also lists our custom voices with a Delete action wired to Hume Delete Custom Voice behind a confirmation, so the library can be pruned.

Two more things to handle. Hume text to speech is rate limited to roughly 100 requests per minute and returns error code E0811 when exceeded, so throttle batch generation, retry E0811 with backoff, and surface which segments failed rather than silently dropping them. And keep the status transitions honest: a script is draft until at least one segment has audio, generated once every segment has a take, and approved only after the publish step succeeds.