Can a prompt replace Podcastle?

Build a local multi-track podcast restoration tool. The goal is making separately recorded speakers sound like they are in the same room.

Project: one track per speaker, imported as separate files, plus optional music and effect tracks. Preserve originals untouched in a read-only folder; every operation is non-destructive and produces a new rendered result.

Analysis first, before any processing. For each track, measure and display: integrated loudness in LUFS, true peak, noise floor during silence, estimated reverberation time, and a spectral average. Show the tracks side by side, because the whole job is closing the gap between them and the user needs to see the gap. Never process a track that does not need it.

Restoration chain per track, each step toggleable with its parameters visible:

1. High-pass filter to remove rumble below the voice range.
2. Noise removal using a noise profile learned from a silent region the user selects, rather than a fixed profile. Apply spectral subtraction conservatively — over-aggressive removal produces the watery artefact that is worse than the noise.
3. Dereverberation, applied only where the measured reverberation time warrants it, with a clear before-and-after preview because this is the step most likely to damage a track.
4. Spectral matching: derive a corrective EQ curve that moves this track's spectral average toward a reference track the user nominates, limited to a maximum of a few dB per band so a bad microphone is improved rather than impersonated.
5. De-esser and a gentle compressor with visible thresholds.
6. Per-track loudness normalisation to a common target before mixing.

Comparison is a first-class control, not a feature: an A/B toggle that switches instantly between processed and original at matched loudness. Matched loudness matters — louder always sounds better, and an unmatched A/B makes every processing chain seem like an improvement.

Assembly: place tracks on a timeline with intro and outro slots, crossfades, and a music bed that ducks under speech automatically. Silence detection with a suggested-cut list the user approves rather than applies automatically.

Export: the mixed episode normalised to -16 LUFS for stereo podcast delivery, each cleaned stem separately, and a project manifest recording every setting so an episode can be re-rendered.

Out of scope: recording remote guests, any voice cloning or synthesis, transcription, collaborative editing, and cloud rendering.