drakulavich/vosk-tts-ru-0.9-multi

Russian Vosk TTS model

Version 0.9

Metrics:

CER 0.6 FAD 0.810 UTMOS 3.290 Speaker Similarity 0.875 xRT CPU 0.35 xRT GPU 0.06

License: Apache 2.0

Changelog:

• —ASR alignment
• —No encoder, just duration predictor
• —Slightly thinner predictor width (160) to fit DiT hidden vector
• —Scale for diffusion loss (to not dominate on duration loss)