Can a prompt replace WellSaid Labs?

Build a local narration studio whose distinguishing feature is per-word control over delivery.

Model: a local TTS engine — Piper or Kokoro — downloaded on first use. Accept only models with a documented licence and record which licence each voice carries, since that field decides whether the output can be used commercially.

Script: paragraphs of text split into sentences, each sentence rendered as an independent audio segment. Segment-level rendering is what makes iteration bearable: fixing one word re-renders one sentence, not the whole script.

Per-word controls, which are the point of this build. Clicking a word in the script opens controls for that word only:

- Pronunciation override, entered as phonemes in the model's own notation, with a preview button that speaks just that word. Provide a phoneme reference and a lookup that proposes a spelling for an unfamiliar word using espeak-ng's rules as a starting point.
- Emphasis: three levels, applied by adjusting the word's duration and pitch contour relative to its neighbours rather than by shouting it.
- Pause after, in milliseconds.
- Speed multiplier for the word, for slowing a number or an acronym.

Store all of this as a per-word annotation layer above the plain text, so the script stays readable and copyable and the annotations can be exported separately.

Project glossary: a pronunciation dictionary at project level, applied to every occurrence. Adding a word from the per-word editor to the glossary is one click, since the same product name will be wrong in every script otherwise.

The review loop, which matters as much as the controls: waveform of the full narration with sentence boundaries marked, click a sentence to play it, and an A/B toggle between the current and previous render of the same sentence at matched loudness. Matched loudness is essential — otherwise every change sounds like an improvement.

Rendering: assemble sentence segments with the declared pauses, normalise each to a consistent loudness before assembly, and normalise the whole track to a broadcast target. Export WAV and MP3 plus a WebVTT caption file generated from the script with the real segment timings.

Provenance: embed metadata recording the model, voice, licence and render date, and note in the README that synthetic narration used publicly should be disclosed.

Refusals: ship no celebrity, public-figure or scraped voices, and do not implement cloning from a sample.

Out of scope: voice cloning, any hosted API, multi-speaker dialogue, video, and translation.