12 days ago
-
Remove links to the private playground repo; note that the PyTorch checkpoints need an unpublished Matcha-TTS patch, ONNX is self-contained
-
ATTRIBUTION: new name, English labels, licence file is LICENSE
-
Renamed to Matcha-TTS-PL-Pretrain: intermediate checkpoint for fine-tuning; link to Matcha-TTS-PL
18 days ago
-
style_map: neutral = 8 (mid-range · normal); 7 is mid-range · slow · question
25 days ago
-
Default synthetic voice = Masiak 1, Staszczyk 0.5, Proszek 0.5 (UTMOS 3.38); female preset with three base-model women at 0.5
-
Five synthetic voice presets (synthetic = default, synthetic2-5)
26 days ago
-
Card: playground temperature default 0.8
-
New default synthetic voice: Bielenia 0.7, Faszczewska 0.2, Masiak 1, Glogowski 0.15, Staszczyk 0.15 (UTMOS 3.49)
27 days ago
-
Card: playground temperature default 0.5
-
Remove stray session file
-
default voice: synthetic (weighted blend of five readers); 7-reader preset removed
28 days ago
-
default voice: Wojciech Masiak; 7-reader blend kept as a preset
-
default voice: 7 readers equal share (Kopa 0), rescaled blending, temperature 0 recommended
-
RECIPE: new voice needs one to two hours of clean recordings
-
default voice: equal blend of all 8 readers (only preset); card and samples updated
-
ONNX with spk_emb input (any blend/style at run time); voices.json = embedding tables + trio preset; Kopa blends removed
-
ONNX: every voice × style baked as speaker rows; default voice trio (Faszczewska+Masiak+Bielenia); structured voices.json
-
v3: Polish Matcha-TTS + fine-tuned HiFi-GAN vocoder, ONNX with baked voices
-
initial commit
