What would larger narration chunks cost against speech policy 6, and at what latency?

Open
#326 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Assessment

Difficulty
5/5
Estimated time
Over a week
Newbie friendliness
15/100
Issue type
Documentation
Clarity
Mostly clear
Activity status
Active
Domain
ai, performance

Research direction

This is an estimation write-up, not a code change. Start with the narration eval fixtures and the derStandard, Anthropic and Cloudflare excerpts, then read the cost and overlap analysis in issue 317 and the chunk-size question in issue 324, since both set inputs. Read ADR 0008 and ADR 0011 for the whole-document synthesis and single-track constraints that define the latency and retry numbers. Done means each figure is labelled measured, estimated or unverified, with the assumptions stated.

Written by the indexing model from the issue text.

Description

wayfinder:research

Question

What would larger narration chunks cost against speech policy 6, and what would they do to the slowest synthesis request and a failed segment's retry?

The baseline is speech policy 6, one request per synchronization unit. Overlap and cut, about 1.45× policy 6 because each segment boundary repeated the previous chunk's last sentence (How much overlap, and at what cost?), was withdrawn on 2026-11-13 (ADR 0028); its cost is now only the price of continuity that larger chunks are measured against. With the chunk size How long can one Gemini 3.8 TTS request get before a fresh start inside it? settles, estimate from the narration eval fixtures (and the derStandard, Anthropic and Cloudflare excerpts):

  • synthesized audio per article against policy 6, plus speech recognition and the model-based placement step per chunk;
  • the price change of Gemini 3.8 Flash-Lite TTS audio on 2027-01-01;
  • the slowest synthesis request per article and the conversion's time until the narration track exists, given that the whole document is synthesized and the track exists only once every segment has audio (ADR 0008, ADR 0011);
  • how much audio a failed segment's retry repeats, and what that does to cost at plausible failure rates.

Mark measured, estimated and unverified figures. Light paid probes are fine.

Part of Map: decouple speech synthesis from synchronization units.

Dominant language
TypeScript
Stars
1
Forks
0
PR merge metrics
No merged PRs in 30d

Getting set up

This project ships no dev container, Dockerfile or contributing guide, so setting up is up to you: start from its README, and see our first-contribution guide for the general steps.

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from patricktree/cup

All issues in patricktree/cup

Similar issues

More TypeScript issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.