Text normalization uses NeMo text processing, which needs pynini 2.1.6.post1. PyPI has no Python 3.14 wheel of it, and no Linux aarch64 or macOS wheel of any pynini version, so take pynini from conda-forge and the package from PyPI:

pixi init e2m && cd e2m
pixi add python=3.14 pynini=2.1.6.post1 editdistance
pixi add --pypi epub-to-m4b
pixi run epub-to-m4b chapters book.epub

$ pixi run epub-to-m4b chapters outliers.epub
Outliers the story of success — Gladwell Malcolm (14 chapters)
  #  title                                               paras    chars  docs
  1  INTRODUCTION — The Roseto Mystery                      27    12584     3
  2  CHAPTER ONE                                            65    26263     8
  3  CHAPTER TWO                                            76    33493     7
  4  THE 10,000-HOUR RULE                                   26     9883     1
  …

Each detected chapter becomes a chapter marker in the .m4b.

• --toc-depth N (default 1): deepest TOC level whose entries start chapters.
• --min-chars N (default 200): chapters with fewer body chars merge into the next one (a trailing stub into the previous).
• dump-text --chapter N --split shows chapter N as the narrator will read it, normalized and split into sentences. --max-chars sets where long sentences are cut (default 125); set it to the engine's limit to see its clips.

• Sentence clips are cached as FLAC under <cache_dir>/clips/<engine-fingerprint>/, shared across books. Assembled chapters are cached under <out_dir>/.work/<book-id>/chapters/.
• Changing engine, voice, model, or other audio settings stops old clips being reused. Old clips stay on disk.
• Clips are keyed by the exact sentence text. A text pipeline change (text/normalize.py, text/split.py, a new nemo-text-processing) re-synthesizes only sentences whose text it changed.
• NeMo compiles a language's grammars on first use into <cache_dir>/nemo/<version>/<lang>/ (E2M_NEMO_CACHE_DIR replaces <cache_dir>/nemo).
• A missing or damaged .m4b, or one encoded with different settings, is rebuilt; an up-to-date one is kept and only the .vtt is rewritten.

A clip is one sentence, or several short ones packed together, so it always ends where a sentence ends. Only a sentence longer than the engine accepts is cut, at a comma where possible: about 125 characters for silence/tone, up to 300 for breeze, 1000 for the API engines, and about 1200 for drama at its default 8192-token context (what fits after the reference voice, in any script).

• Breeze: batch_size is clamped to the server's reported limit. The server runs one inference at a time; a second client waits for it.
• Drama: a cold start takes about ten minutes. The reference voice is uploaded once, named after its audio and transcript, so a changed file or transcript uploads a new voice. batch_size sentences are sent at once and two batches run together; the server decodes them in one GPU batch.