A 1.1-second fixed overhead pushes caption-driven TTS out of the conversational path

The baseline is the only absolute number worth anchoring on: 0.8 seconds of generation for a 7.5-second sentence ^[1][4], an RTF of roughly 0.11 ^[18]. Apply the two measured ratios to that and the diffusion model needs about 2.0 seconds at 40 steps and about 1.2 seconds at 16 ^[19][20]. The deciding figure sits underneath both, a fixed overhead of about 1.1 seconds that does not move when you cut steps or add compute ^[8]. That is roughly 1.4x the incumbent's entire generation time, spent before a single denoising step runs ^[21]. At the 12-step floor of 1.1 to 1.4 seconds, overhead accounts for 79 to 100 percent of the clock ^[8][22].

The writeup attributes that overhead to model loading and text processing, neither of which benefits from parallelism ^[8]. Those two behave differently inside a server. Model load is paid once per process and amortizes across every request a warm worker serves. The Japanese front-end work is per request and stays. The measurement does not report how the 1.1 seconds divides between them, and that division decides whether a resident worker recovers most of it or none of it. The method the author recommends, cutting steps until generation stops improving ^[9], is the right instrument for sizing the parallelizable share; I would run it twice, cold and warm, because the difference isolates load cost from per-request cost without touching the model.

The 2.5x is a claim about that harness ^[5]. One diffusion implementation, one 7.5-second Japanese sentence, one 2g.20gb slice ^[4][7], measured against a Style-Bert-VITS2 model the team had already trained ^[3]. For the ratio to transfer you need the same playback shape: synthesis and playback sentence by sentence, first-sentence delay as the entire budget, and a 1-to-2-second turn treated as unusable ^[6]. Buffer a whole reply before playing it and the fixed cost spreads over a longer clip. Move to a smaller slice and the gap widens, which the source states outright ^[23].

The design ledger is the part I would carry to any stack. Caption plus seed is deterministic, so the reproducible input is the asset and the trained model file is a derivative ^[2][13]. The author lost a driver script and rebuilt the voice from a caption and seed recovered out of logs ^[13]. Not a restore drill anyone schedules.

The split reads as free only if you already own the runtime half. If you do not, adopting it buys a training pipeline, a 200-clip corpus per character, and a quality ceiling set by the generator, since its output becomes the training data ^[11][14]. Speaker identity has to hold across those 200 clips, and exclamation-initial lines broke it until reference audio was managed by hand ^[15]. Where playback is not interactive, the 1.5x gap at 16 steps is affordable ^[5], and the caption model can stay in the path.