The check costs two calls and a subtraction. Set the multiplier to the extremes of its documented range, measure the returned file, compare the delivered ratio to the ratio you asked for.
On the published numbers, 0.85 and 1.35 command a duration ratio of 1.59, so the two clips should differ by about 59 percent [15]. They differed by 2.5 percent [16]. That is roughly four percent of the commanded change [17], and the shape is wrong as well: the 0.85 request, the faster one, came back a tenth of a second longer than the call that passed no parameter at all [18]. Read as a plain multiplier, the 1.35 clip landed about 1.2 seconds short of where it should have [20].
Four calls, one text, one style, no repeats [3]. The author reads the 5.1 to 5.3 spread as ordinary synthesis fluctuation [5]. With a single sample per setting you cannot separate fluctuation from a very weak real effect, and I would want repeats before writing that sentence, though neither reading rescues a parameter that delivers four percent of its range.
The response layer said nothing about any of this. The API accepted the field, returned no error, and produced normal audio [6]. The team's diagnosis is that the server side never forwarded the value to the synthesizer [6]. Their harness computed mora count over measured seconds of the returned waveform [3], which is the only oracle that could have caught it. A contract test asserting a 200 and a non-empty body passes here, every time.
The post does not name the API or the model [21], so it cannot tell you your provider has this bug. What transfers is the method, and the prior behind it: a request field that shapes output, rather than selecting a resource, is untested until you measure the output. The author's version of that rule is that "can pass" and "takes effect" are different things [7].
Because prosody could not be moved at synthesis time, it moved into corpus selection [8]. A conv_style value now picks both a script set and the strictness of the quality gate, across 14 profiles [9]. Inside build_corpus_plan, the per-category count of extreme acting material comes from EXTREME_PER_CAT with a default of 8, and is forced to 0 for the 11 profiles in the business set [11][23]. Business voices end up with 12 styles, casual, mixed and mc with 17 [12], the difference being the five acting styles [19].
The last link is the part worth copying. The runtime's enthusiastic reaction fires on the presence of "Scream" in a model's style list [13]. A model trained without screaming material has no such entry, so the feature has nothing to trigger on, and the stated goal was accident prevention rather than audio quality [14]. That is a guardrail made of a missing row rather than a flag someone can flip at three in the morning.
The bill is that "make the support voice slower" is now a corpus edit and a training run [8]. In a system whose worst output is a call centre line laughing at a customer [14], I would take that trade as well.
