Reply to thread

Message: [QUOTE="Bombastus, post: 91169, member: 2178"] ACX currently requires human narration unless otherwise authorized, even when an AI-generated file passes every technical audio measurement. Eligibility and file quality are separate review decisions, and confusing them wastes mastering time. A convincing voice sample proves very little about an eight-hour production. Quality control must establish that the complete text is present, the performance remains stable, and every delivered file meets the chosen retailer's requirements. That work belongs inside [B][URL='https://goldmidi.com/community/threads/what-ai-narration-did-to-audiobook-production-costs.76576/']the hidden cost of synthetic narration[/URL][/B] because cheap generation can still demand hours of structured proofing. The fastest process catches language defects before anyone spends time mastering the waveform. [HEADING=2]Text fidelity fails before mastering begins[/HEADING] Freeze the source manuscript before generating audio. Create a chapter manifest containing each title, its opening and closing words, and any material intentionally excluded from narration. Without that reference, a shortened chapter can look complete merely because the file exported successfully. Mark source elements that should remain unheard, including page headers, footnote markers, image captions, or navigation text. Silence is not always an omission, and an exclusion record prevents a proofer from restoring material the audio deliberately leaves out. Run [B]a manuscript-to-audio error pass[/B] while following the text on screen. Mark omitted, repeated, altered, or invented words with timestamps, including tiny changes that reverse meaning. A casual listen rarely catches every missing article, negative, or attribution. Build a separate inventory of risky language. Include character names, place names, homographs, dates, decimal numbers, abbreviations, web addresses, formulas, foreign phrases, and words with disputed pronunciations. Search for every occurrence after making a correction, since fixing one rendition does not guarantee that later chapters use it. Write the intended spoken form beside ambiguous numbers and abbreviations. A year can be read as a quantity, while the same initialism may need letters in one passage and a word in another. This converts a subjective correction into a repeatable acceptance test. Chapter boundaries need their own inspection. Confirm the final sentence of one chapter, the first sentence of the next, the displayed chapter name, and the actual playback order. File numbering helps organization, but it cannot prove that an export contains the intended material. [HEADING=2]Human listening catches what meters cannot[/HEADING] Waveform measurements cannot judge whether a voice changes age halfway through a book or delivers grief with cheerful energy. A listener must also catch rushed clauses, unnatural pauses, unstable accents, and audible joins around regenerated passages. These defects can survive perfect text alignment. Use two listening modes for different jobs. First, follow the manuscript closely and verify every spoken word. Then listen without reading on headphones and an ordinary speaker, which exposes pacing and continuity problems that visual concentration can hide. Choose stress passages deliberately instead of relying on random spot checks. Include rapid dialogue, a long sentence, a list, foreign vocabulary, an emotional reversal, and the final chapter, where fatigue or model drift may be most obvious. Keep a correction log with the timestamp, source sentence, defect, replacement, and approval status. Regenerate the smallest workable segment, then listen across the edit from several seconds before it to several seconds after it. A corrected word can still leave a sudden change in pace, volume, or vocal character. Repeated terms deserve a final consistency check after all pickups are complete. A name pronounced correctly in chapter one but differently in chapter twelve sounds careless, even when both versions are individually intelligible. Searchable timestamps turn that check into a finite task instead of another full listen. [HEADING=2]Technical compliance belongs at the end[/HEADING] Mastering unfinished narration creates avoidable rework. Every regenerated segment can change timing, loudness, silence, or peak level, so lock the wording and performance before normalizing the complete set. Measure the exported delivery files rather than an earlier working copy. For an authorized ACX production, its analysis tool checks measurements including RMS, peak levels, and sample rate against the submission requirements. Kobo applies a different packaging test, accepting MP3 files while limiting each file to 200 MB and the complete upload to 2 GB. Passing one retailer's validator does not establish compliance everywhere. Compare chapters as a batch after checking them individually. Two files can satisfy the permitted measurement range yet create an obvious volume jump in continuous playback, which remains a listener problem even when the validator reports success. Run a package audit after encoding. Compare the delivery folder with the chapter manifest, confirm file order and duration, and verify the opening credits, closing credits, retail sample, cover, and metadata. Listen to the first and last seconds of every file because clipping and accidental carryover often sit at boundaries. Keep the frozen manuscript, raw generations, approved edits, mastered files, correction log, and retailer exports as separate versions. When a store reports a defect, that record identifies whether the problem came from narration, mastering, packaging, or the wrong upload. [/QUOTE]

Name