30 days ago
-
Fix: Peak audio normalization (-1 dB), robust speaker search, and tuned CFG for crystal-clear benchmark generation
-
Fix: Peak audio normalization (-1 dB), robust speaker search, and tuned CFG for crystal-clear benchmark generation
-
Fix: Eliminate 30MB raw audio waveform IPC serialization to permanently prevent Docker /dev/shm shared memory crashes
-
Turbo Speed: Vectorized FFT F0 pitch, cuDNN autotune, Tensor Core precision, speaker RAM cache & prefetch DataLoader
-
Turbo Speed: Vectorized FFT F0 pitch, cuDNN autotune, Tensor Core precision, speaker RAM cache & prefetch DataLoader
-
Master Fix: Real F0 pitch, acoustic Mel-80 to 100 filterbank vocoder projection, soundfile multi-format loader, safe weights_only
-
Fix: CUDA OOM prevention via expandable_segments, audio duration clamping, empty_cache, and safe 3090 batching
-
Fix: Docker container /dev/shm exhaustion by setting mp sharing strategy to file_system and num_workers to 2
-
Fix: check hasattr for sampler.set_epoch on WeightedRandomSampler
-
Performance: 32 parallel download workers and hf_transfer for 10x faster dataset downloading
-
Performance: 32 parallel download workers and hf_transfer for 10x faster dataset downloading
-
Cloud-Sync & Auto-Resume: Rental GPU data protection with --push_to_hub and automatic cloud recovery
-
Master SOTA Upgrade: in-context voice conditioning, decoupled studio vocoder, EMA tracking, warmup scheduler
-
Clean train.py: save local step checkpoint only, no auto-upload
-
Live auto-sync every 10k step milestone directly to HF
-
Add --save_every_steps and max_steps instant checkpoint saving
-
Update: auto-download HF dataset fallback in train.py
-
Update: training fixes, LICENSE + requirements + scripts
-
Initial commit: ViuAI_TTS_200M architecture, prosody engine, and weights
-
Initial commit: ViuAI_TTS_200M architecture, prosody engine, and weights
