VITS / END-TO-END TTS / 2021

VITS · LJSpeech

TOTAL≈36Mparameters

ACTIVE≈36Mruntime path

DEPTH6 + flowrepeated blocks

HIDDEN192main trunk

CONTEXTutterancestream / sequence

CHECKPOINT≈145MB1 shards · MIT

INTERACTIVE SYSTEM BLUEPRINT

MODEL MRI

128 tokens → 10 seconds · 70.51G arithmetic FLOPs · click any module for evidence

VERIFIED ADAPTER

INPUT128TOKENS

TRANSFORMER STACKVAE + FLOW + GANtext prior ↔ posterior latent → parallel waveform decoder

Multi-Module Residual

OUTPUT10SECONDS

wraps every repeated block

TEXT SELF-ATTENTION · STOCHASTIC DURATION PREDICTOR时长不是确定回归，而是从条件分布采样，保留韵律多样性。

Normalizing Flow VAE训练用 posterior 声学信息，推理把文本 prior 通过逆 flow 映射到 decoder latent。

TEXT6LTransformer

HIDDEN192encoder

FLOW4coupling blocks

VOCODERGANdecoder

SAMPLE22.05kHz

PARAMS≈36MLJS