Sampling block for upsampling or downsampling

Vocos backbone module built with ConvNeXt blocks. Supports additional conditioning with Adaptive Layer Normalization Args: input_cha

BiCodec tokenizer for handling audio input and tokenization.

Decoder module with convnext and upsampling blocks Args: sample_ratios (List[int]): sample ratios example: [2, 2] means downs

Follows Algorithm 1. in https://arxiv.org/pdf/2107.03312.pdf

Adaptive Layer Normalization module with learnable embeddings per `num_embeddings` classes Args: condition_dim (int): Dimension of t

Encoder module with convnext and downsampling blocks

Multi head attentive statistics pooling Reference: Self Multi-Head Attention for Speaker Recognition https://arxiv.org/pdf/1906.0

An attentive pooling Reference: multi query multi head attentive statistics pooling https://arxiv.org/pdf/2110.05042.pdf Args

Spark-TTS for text-to-speech generation.

Args: input_dim (int): acoustic feature dimension out_dim (int): output dimension of x-vector and d-vector latent_dim (i

ConvNeXt Block adapted from https://github.com/facebookresearch/ConvNeXt to 1D audio signal. Args: dim (int): Number of input channels.

in_channels == out_channels == channels

ResBlock adapted from HiFi-GAN V1 (https://github.com/jik876/hifi-gan) with dilated 1D convolutions, but without upsampling layers. Args

Attentive statistics pooling: Channel- and context-dependent statistics pooling, first used in ECAPA_TDNN.

Base class for the generator's backbone. It preserves the same temporal resolution across all layers.

BiCodec model for speech synthesis, incorporating a speaker encoder, feature encoder/decoder, quantizer, and wave generator.

Temporal average pooling, only first-order mean is considered

Temporal standard deviation pooling, only second-order std is considered

Temporal statistics pooling, concatenate mean and std, which is used in x-vector Comment: simple concatenation can not make full use of b

Turn label to special token

Triton Python model for Spark TTS. This model orchestrates the end-to-end TTS pipeline by coordinating between audio tokenizer, LLM, and

Triton Python model for audio tokenization. This model takes reference audio input and extracts semantic and global tokens using BiCodec

Triton Python model for vocoder. This model takes global and semantic tokens as input and generates audio waveforms using the BiCodec voc

Vocos backbone module built with ResBlocks. Args: input_channels (int): Number of input features channels. dim (int): Hidden