Whisper transcribes a candidate clip, a judge compares that transcript against the script, and any deviation fails the clip [2]. Line that up against what each emotion does to the waveform. Fear comes out as a tremble, joy extends or raises the pitch at the end, anger distorts the volume, sadness fades at the finish [6]. Every one of those makes recognition harder, so the transcript drifts and the take is discarded, while the flat read stays legible and passes [6]. The original loop returned as soon as a clip passed, so seed order combined with Whisper's preference chose the corpus, and the flattest surviving take was reliably first in line [11]. Nothing in that path threw an error; the pipeline was working exactly as configured [16].
The ear did not catch it either. The developer reports thinking the samples were fine on listening, and only reconsidered when cosine similarity to neutral came back near 0.9 [15]. At that value the joy style is emitting something close to the neutral voice, which makes the style controls decorative [14].
Convert those similarities into the distance the fix actually ranks on, 1 minus cosine [8]. The bulk corpus lands between 0.06 and 0.23 from neutral; the hand-made clip lands at 0.836 [1]. The good clip therefore sits about 3.6 times further from neutral than the best bulk clip, and roughly 14 times further than the worst [2]. A pass/fail gate never measures that gap; it only checks whether the transcript matches exactly.
Underneath the selection bug were two separate configuration faults [18]. Scripts that opened with an interjection such as "Waa!" made the start of the clip sound like a different speaker, because the parameter governing fidelity to the reference audio broke speaker consistency once emotion was applied [12]. Emoji used as inline emotion markers were read aloud as "happy face" or triggered sound effects, which is what happens when markup shares a channel with content [13].
The fix has a bill. The old loop stopped at the first passing clip, and the new one collects every candidate before ranking, so each line now pays synthesis plus Whisper transcription for the full seed count [3]. If no candidate passes, the line yields nothing at all [7].
On transfer: these cosine figures come from one model's own style embedding space, one speaker, and a single hand-made reference clip rather than a distribution [5][8]. Read 0.164 as an existence proof that the styles can separate under some settings, not as a target for your speaker. The portable part is the shape of the error. When the pass criterion of a filter correlates with the property you are trying to maximise, ranking by pass order optimises against you.
