EP 004 — It Passed QA. It Still Wasn't Me.
from AI with Pie Hosted by Pie
What happens when an AI-generated voice says the right words, passes technical QA — and still does not sound consistently like the person it is meant to represent? In Episode 004, Pie and her AI co-host revisit a real AI-with-Pie production case. A short voice test sounded convincing. The longer episode did not. Some passages felt warm and recognizably Pie; others sounded younger, more metallic, more distant or simply unlike the way she hears herself speaking English. The difficult part was not finding one perfect setting. It was learning what actually needed approval: not a voice label or configuration, but each concrete take, sentence, thought group and assembled episode. They discuss: - why a technically clean file can still fail a human identity gate; - why repeated regeneration is not automatically an improvement; - how local repair protects performances that are already right; - why content, technical and identity checks answer different questions; - why the complete assembled episode still requires a full human listen; - three practical rules for working with synthetic media that represents a real person or brand. The next case begins with one tiny visual change — and a different room. AI transparency AI-generated voices. Human judgment. Pie's voice is generated from her own recordings with her consent; the AI co-host uses an AI-generated synthetic voice. This authored conversation was developed with AI, then reviewed, shaped and approved by Pie. It is an editorial reconstruction, not an unedited recording of a live conversation. Case and evidence note This episode reconstructs a documented AI-with-Pie production case. It reports observed voice inconsistency and the resulting repair process without claiming one proven technical cause. It does not identify clients, participants or private third parties and does not include private historical before-and-after voice clips. Artwork transparency The Show Cover includes AI-assisted imagery, art-directed and approved by Pie. Chapters (00:00) AI-generated voices. Deliberate human direction. (00:12) Episode highlight: it passed QA (00:21) The tinny voice from Oz (01:29) A real-world field test (01:51) The part that made no sense at first (02:58) The setup looked the same. The result didn't. (03:59) The tempting fix: render it again (04:30) Why micro-patches were not automatically better (04:55) Then technical QA passed (05:53) The real diagnosis (06:30) Repair the affected layer (06:49) Why the full listen changed everything (07:59) The rerun (08:48) Why this is bigger than a podcast voice (09:50) Three Things to Go (10:56) The core shift: stabilize the approval process (11:36) Next case: one tiny change, a different room (12:10) AI with Pie. Useful. Awkward. Fun. Website https://aiwithpie.com
Embed this episode
No transcript yet.