How to Batch-Produce Audio Chapters with the Listnr API

November 16, 2026

Turning a manuscript into audio by hand means pasting text and downloading files a dozen times. The Listnr AI API removes that loop, but it introduces a different problem: long text blocks time out, and returned files arrive out of order.

This is the production pattern that handles both.

How does the Listnr API work?

The API accepts text plus a voice identifier and returns an audio URL or a job handle depending on the endpoint version you are on. Authentication uses a bearer token from your account settings.

Two practical constraints shape everything. First, request timeouts: sending a 30,000-character chapter in one call is a common way to get a gateway error. Keep requests under roughly 5,000 characters. Second, returned audio is segmented, so a long chapter comes back as multiple files or as one file with slightly different pacing at segment boundaries.

Before writing any code, send one 500-character request manually and inspect the response shape. Endpoint fields have changed between versions, and guessing wastes an evening.

How do you split a manuscript for reliable rendering?

Split at paragraph boundaries, then merge paragraphs into chunks of 3,000 to 4,000 characters. Do not split mid-sentence — every split point is a place where the voice resets its intonation, and a reset mid-clause sounds broken.

A workable pipeline:

  1. Parse the manuscript into chapters, then into paragraphs.
  2. Accumulate paragraphs into chunks under 4,000 characters, breaking only between paragraphs.
  3. Submit chunks in parallel with a concurrency limit of three or four.
  4. Store chunk index alongside each returned file.
  5. Concatenate in index order, then listen to every join point.

Parallelism matters for turnaround, but unbounded concurrency gets you rate limited and half your job fails. Four simultaneous requests is a safe default unless your plan documents otherwise.

A 90,000-word manuscript becomes roughly 600,000 characters, or about 150 chunks. At a few seconds each, serial processing takes an hour; at four-way concurrency it finishes in under twenty minutes.

How do you handle failures and retries?

Retry on timeout and 5xx with exponential backoff — start at two seconds, double, cap at thirty. Do not retry on 4xx; a bad voice ID will fail every time.

Log the chunk index, character count, and response ID per attempt. When a retry succeeds, delete the earlier partial file, or your concatenation step silently doubles a paragraph.

Validate character counts before submitting. A chunk that is mostly markup or contains stray UTF-8 from a PDF conversion can produce audio that reads out formatting artefacts. Strip asterisks, brackets, and footnote markers first — the model will otherwise pronounce them.

How do you stitch and master the result?

Use ffmpeg to concatenate the chunk files, then normalise loudness in the same pass. A single command handles both, and loudness normalisation is what makes the joins inaudible — differing segment levels are the giveaway that a file was assembled from pieces.

After stitching, add a small gap between chapters: 800ms to 1.5s reads as natural, and anything under half a second feels rushed at a chapter break.

Chapter headers matter for distribution. Most audiobook platforms require per-chapter files with metadata, not one long file. Export chapters separately even if you also keep a stitched master for review.

What are the costs and limits?

Everything is metered in characters. A 90,000-word manuscript is over half a million characters, which will exceed small plans in one book. Check the per-character rate and whether retries are billed — some providers charge for failed renders, which makes handling timeouts carefully a financial concern, not just an engineering one.

Voice consistency across a long manuscript is decent but not perfect. Render the whole book with the same voice ID in one session; switching voice versions mid-project produces audible drift at the seam.

And there is no automatic proofing. A mispronounced name that appears 200 times appears wrong 200 times, so test the first chapter and listen closely before rendering the rest.

Quick Summary

  • Keep API requests under about 5,000 characters; long blocks time out.
  • Split at paragraph boundaries and accumulate into 3,000 to 4,000 character chunks.
  • Run three to four requests in parallel, not unbounded concurrency.
  • Concatenate with ffmpeg and normalise loudness so chunk joins are inaudible.
  • Budget by characters: a 90,000-word book is roughly 600,000 characters.