Local Kokoro Long-Form Narration — Run 001

Status · SUPPORTED

Research class

RESEARCH REPORT

Completion date

July 27, 2026

Research question

Can the local Kokoro production lane render and deliver one complete long-form N’OMOTO article without a speech API request?

Prior hypothesis

A lightweight local Kokoro model could render a complete single-narrator article with resumable paragraph units, exact source lineage, deliberate paragraph pauses, and standard web playback.

Method

One published 2,872-word Maimonides article with no existing audio was divided into 24 authoritative paragraph units after excluding one obsolete promotional opening. Each unit was rendered locally with the Kokoro af_heart voice at a restrained rate, preserved with text and audio hashes, assembled into one MP3, fully decoded, uploaded, range-tested, attached to the canonical article, and played through the public browser interface before and after refresh.

Observations

• All 24 paragraph units rendered locally and remained resumable.
• The complete narration contains 2,872 source words and runs for approximately 18 minutes 36 seconds.
• The assembled MP3 is non-empty, decodable, and served with HTTP partial-content support.
• The public article exposes Read and Listen without a YouTube action for that work.
• Browser playback started, paused, and started again after refresh.
• The obsolete Selenius Media promotional opening was excluded while the substantive article text remained the narration source.

Supporting evidence

• The source and narration hashes are recorded and the 24 segment indexes are sequential.
• The final MP3 hash matches the uploaded WordPress media file.
• Full-file decoding completed without an audio error.
• The public media route returned HTTP 206 for a byte-range request.
• The browser reported a non-zero 1,116-second duration and ready playback state.
• The canonical public article returns HTTP 200 and links the verified audio asset.

Counterevidence

• The owner has not yet listened to the entire 18-minute result.
• Only one reflective educational work has been rendered long-form.
• No dramatic short story has completed the work-specific comparison.
• No multi-speaker, cloning, speech-to-speech, dubbing, or translation capability was tested.
• Kokoro provides less direct expressive prompting than Eleven v3 or the current OpenAI renderer.
• This run establishes technical production and delivery, not equality with every commercial voice system.

Conclusion

Local Kokoro can complete the technical single-narrator Article → Listen workflow for one 2,872-word work without a speech API request. Long-form artistic adoption remains conditional on full human listening, and this result does not establish ElevenLabs-equivalent expressive, cloning, dialogue, or dubbing capabilities.

Confidence

HIGH for complete local rendering, delivery, and browser playback; LOW for long-form artistic preference and untested commercial-provider capabilities

Limits

• One educational article cannot establish a general production default.
• Complete-owner listening remains pending.
• The work used one voice and one narration profile.
• The proof did not test dramatic acting, dialogue, cloning, dubbing, or multilingual production.

Production consequence

The local Kokoro lane is now eligible for bounded single-narrator article tests. Azure remains a parallel credit-backed production lane. Batch replacement and the short-story profile remain on hold until full listening and a dramatic-work proof pass.

Next unresolved question

Does a complete dramatic short story require a different Kokoro voice or pacing profile, and does the owner prefer both work-specific profiles after full-length listening?