## v1.1.004 — 2026-05-06
- fix: voice clone quality on imported reference audio — some end users reported that voices cloned from uploaded files (MP3 / M4A / OGG / FLAC, or non-standard WAV variants) sounded like the wrong speaker, with cases of the cloned voice's perceived gender flipping vs the source. Root cause was that the voice engine was loading the on-disk reference file at whatever sample rate / channel layout / encoding the user had supplied, and the speaker-embedding extractor's behavior on unusual layouts varied subtly across machines (different audio backend builds, different downstream resampler implementations). All newly saved reference clips are now normalized to a single canonical layout (24 kHz mono 16-bit PCM WAV) on save, regardless of the input container. The conversion runs through the bundled FFmpeg, so MP3 / M4A / OGG / FLAC / WebM-Opus uploads are all supported. Reference clips saved before this version are unaffected.
- fix: voice clone audio had audible periodic distortion on the Vulkan voice engine — long sentences contained a dropout / "smear" roughly every second. Root cause was that the vocoder ran in an isolated worker process (introduced in v1.1.002 to fix a different stability issue), and that worker decoded each piece of the codec stream as an independent one-shot graph; the 5 ms equal-power cross-fade we used to bridge the pieces produced a 3–6 dB envelope notch at every boundary because the two adjacent pieces were correlated audio rather than uncorrelated streams. The worker now maintains a true streaming session across pieces — transformer key/value cache, causal-convolution tail buffers, and transposed-convolution overlap-add state are all carried forward — so concatenation is mathematically equivalent to a single-shot decode of the full sequence. The cross-fade has been removed from this path. Smoke-tested at cosine similarity 0.999986 vs single-shot reference decode.
- fix: voice synthesis instability on RTX 30-series cards — the partial workaround shipped earlier (which kept the standard cooperative-matrix path enabled and only disabled the vendor-specific variant) reproduced rc=11 SEH crashes on RTX 3060 Ti within the first few requests after warmup, with failure rates climbing to 80–88 % before auto-respawn. Both cooperative-matrix paths are now disabled on all Vulkan-running GPUs. Net throughput cost of the disable is small relative to the stability win.
- change: long-sentence cap on the Vulkan voice engine raised back to ~81 s of audio. An earlier reduction to ~41 s was a precaution against a runaway-decode failure mode, but it truncated long-sentence synthesis for some end users. Runaway worst-case is still bounded by the in-engine attractor detector that watches the decode loop and aborts within ~5 s if the model gets stuck, so the cap can safely be looser.
- fix: the Voice Clone window's drop-zone hint and oversize-file error message now show the correct 30 MB reference-audio limit. The strings still said "10 MB" across all 30 UI languages, mismatching the actual backend limit (which was raised to 30 MB in v1.1.003).
- improvement: voice quality on long synthesis is slightly cleaner — the Vulkan voice engine's matrix-multiply path now uses 32-bit accumulation throughout the vocoder, eliminating cumulative numeric drift that produced occasional sample-level glitches on outputs longer than ~10 seconds. Cosine similarity vs the CPU reference improved from ~0.9989 to ~0.9998 on 16-second test clips.
- diagnostics: voice synthesis now logs the active reference-audio file path and size once per synthesis (helps remote troubleshooting when a user reports a wrong-voice symptom). When the TTS manager has to fall back to "first available voice" because the requested voice id can't be resolved, this is now flagged with an explicit marker in the log. On voice-engine startup, the resolved profile directory and the counts of cloned / built-in voices are printed once.