## v1.1.003 — 2026-05-05
- fix: voice clone quality — cloned voices sounded distorted, with occasional pops and noise. Two independent root causes: (1) the post-synthesis audio pipeline applied an aggressive dynamic-range compressor (5 ms attack, 2.5:1 ratio) followed by a tanh saturation stage to every output. That pipeline was tuned for the older voice engines whose raw output needed level shaping; it was destructive when applied to the current voice engines whose output is already well-leveled — it pumped on consonant transients (audible as noise) and added odd-order harmonic distortion on vowel sustains (audible as a "distorted" timbre, especially noticeable on cloned voices that preserve the original speaker's chest resonance). The pipeline now has 'standard' and 'gentle' modes; current engines route to gentle (DC-block + 80 Hz high-pass + safety limiter at 0.95 + 30 ms fade only). (2) When the streaming vocoder session was unavailable and decode fell back to independent 480 ms chunks, the chunks were concatenated with a hard butt-join, producing a click at every boundary on long sentences. Fallback paths now use a 5 ms equal-power cross-fade.
- fix: cloned voices created in older sessions stopped showing up after switching to the new voice engine. The newer engine builds were pointing at their own per-engine profile directories instead of reusing the original on-disk store. All current voice clone engines now share a single profile directory, and a one-time migration on first launch merges profiles, design presets and style instructions from any legacy per-engine directories. Old directories are kept as cold backups with a marker file to prevent re-migration.
- fix: the Fast Voice clone window showed "loading failed" in the My Voices section and could not save or delete voice profiles. The IPC handler suite for the shared on-disk profile store had been disabled together with the legacy voice engine in v0.97, but Fast Voice clone still depends on those handlers. Handler suite is restored; the dormant legacy engine instance no longer spawns its underlying service.
- change: Fast Voice clone reference-audio length limit raised from 3–10 s to 3–30 s, matching the High-Quality Voice clone limit. The 10 s ceiling was inherited from the legacy engine pipeline and did not apply to the Fast Voice model; users uploading 12–15 s reference clips were being rejected.
- News and updates
- AI Desktop Pet
- v1.1.003 — 2026-05-05
Source: Steam Community (opens in a new tab)
