A self-contained utility module that finds an ffmpeg binary (preferring system install, falling back to imageio-ffmpeg, then environment variable) and provides a robust `get_duration_video` function for any audio/video file. It first attempts soundfile for fast WAV duration lookup, then falls back to parsing ffmpeg stderr output. Only 3 external imports (os, re, shutil, subprocess, soundfile, imageio-ffmpeg — all standard or thin wrappers). Could be extracted into a standalone package for any Python project needing ffmpeg discovery and media duration probing.

A lightweight prompt loader that reads YAML prompt files from a directory, caches them with `lru_cache`, and fills `{placeholder}` values using `str.format_map` with a custom `SafeDict` that leaves unfilled placeholders unchanged rather than raising `KeyError`. This is a clean, reusable pattern for managing LLM prompts as external files — avoids embedding templates in code, handles missing variables gracefully, and could be extracted with zero modifications (only depends on `pathlib` and `yaml`).

A `CostTracker` dataclass that tracks wall-clock times per pipeline step (recording elapsed seconds and step name), accumulates token/character usage from LLM and TTS calls with a thread-safe lock, and produces a serializable `cost_breakdown()` dict that computes estimated dollar costs using external pricing from `pipeline/pricing.py`. The `print_summary()` method outputs a formatted cost/time table. This is a reusable telemetry component for any multi-step API pipeline — it is self-contained, relies only on a simple pricing module, and uses a threading lock for concurrency safety. Only dependency is the `pricing` and `config` modules, which could be injected or replaced for extraction.

Provides the `Segment` dataclass and a suite of functions for cleaning and structuring Whisper transcription output: `merge_continuous_segments` deduplicates overlapping fragments and merges non-sentence-ending segments for natural TTS input; `split_into_sentences` performs sentence-level segmentation with time distribution proportional to character count and soft-splits long sub-sentences at clause punctuation; `_deduplicate_fragment` strips overlapping prefix text between adjacent Whisper chunks; and `_soft_split_long` recursively splits at the clause punctuation closest to the midpoint. The hallucination filter and abbreviation whitelist demonstrate thoughtful real-world tuning. These data processing functions are well-isolated in the module and could be reused in any project working with Whisper or word-level timestamps.