A modified Snake function which uses separate parameters for the magnitude of the periodic components Shape: - Input: (B, C, T)

A modified Snake function which uses separate parameters for the magnitude of the periodic components Shape: - Input: (B, C, T)

r""" This is the configuration class to store the configuration of a [`Qwen3TTSTalkerCodePredictorModel`]. It is used to instantiate a Qwen3TT

r""" loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided): Language modeling loss (for next-token pr

Layer scale from [Touvron et al 2021] (https://huggingface.co/papers/2103.17239). This rescales diagonally the residual outputs close to 0, with a

Container for one sample's voice-clone prompt information that can be fed to the model. Fields are aligned with `Qwen3TTSForConditionalGener

This class implements an attentive statistic pooling layer for each channel. It returns the concatenated mean and std of the input tensor.

This class implements an attentive statistic pooling layer for each channel. It returns the concatenated mean and std of the input tensor.

Efficient distributed group residual vector quantization implementation. Follows Algorithm 1. in https://arxiv.org/abs/2305.02765 Group Then r

Efficient distributed residual vector quantization implementation. Follows Algorithm 1. in https://arxiv.org/pdf/2107.03312.pdf

An implementation of the speaker embedding model in a paper. "ECAPA-TDNN: Emphasized Channel Attention, Propagation and Aggregation in TDNN Ba

Codebook with Euclidean distance. Args: dim (int): Dimension. codebook_size (int): Codebook size. kmeans_init (bool): Whet

Calculate the BigVGAN style mel spectrogram of an input signal. Args: filter_length (int): The number of samples in the filter window

Multi-headed attention from 'Attention Is All You Need' paper

An implementation of the speaker embedding model in a paper. "ECAPA-TDNN: Emphasized Channel Attention, Propagation and Aggregation in TDNN Ba

r""" This is the configuration class to store the configuration of a [`Qwen3TTSSpeakerEncoder`]. It is used to instantiate a Qwen3TTS speaker

Multi-headed attention from 'Attention Is All You Need' paper

r""" This is the configuration class to store the configuration of a [`Qwen3TTSTalkerModel`]. It is used to instantiate a Qwen3TTSTalker model

r""" loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided): Language modeling loss (for next-token pr

r""" This is the configuration class to store the configuration of the Qwen3TTSTokenizerV1DecoderToken2WavBigVGAN module. It defines the archi

r""" This is the configuration class to store the configuration of a [`Qwen3TTSTokenizerV1DecoderConfig`]. Configuration objects inherit from

r""" This is the configuration class to store the configuration of the Qwen3TTSTokenizerV1DecoderToken2WavDiT. It defines the architecture of

r""" audio_values (`List[torch.FloatTensor]`): Decoded audio values, obtained using the decoder part of Qwen3TTSTokenizerV1. Each

r""" This is the configuration class to store the configuration of the Qwen3TTSTokenizerV1 Encoder. The encoder typically takes mel-spectrogr

r""" audio_codes (`List[torch.LongTensor]`): Discret code embeddings computed using `model.encode`, each tensor has shape (codes_length_i,

Multi-headed attention from 'Attention Is All You Need' paper

r""" This is the configuration class to store the configuration of a [`Qwen3TTSTokenizerV2DecoderConfig`]. Configuration objects inherit from

r""" audio_values (`List[torch.FloatTensor]`): Decoded audio values, obtained using the decoder part of Qwen3TTSTokenizerV1. Each

r""" audio_codes (`List[torch.LongTensor]`): Discret code embeddings computed using `model.encode`, each tensor has shape (codes_length_i,

Residual Vector Quantizer with separate projections for the first quantizer and the rest. Args: n_q (int): Number of residual vector quan

An implementation of building block in ECAPA-TDNN, i.e., TDNN-Res2Net-TDNN-SqueezeExcitationBlock.

An implementation of building block in ECAPA-TDNN, i.e., TDNN-Res2Net-TDNN-SqueezeExcitationBlock.

Vector quantization implementation. Currently, supports only euclidean distance. Args: dim (int): Dimension codebook_size (int

This is the configuration class to store the configuration of a [`Qwen3TTSForConditionalGeneration`].

A HuggingFace-style wrapper for Qwen3 TTS models (CustomVoice/VoiceDesign/Base) that provides: - from_pretrained() initialization via AutoM

r""" Constructs a Qwen3TTS processor. Args: tokenizer ([`Qwen2TokenizerFast`], *optional*): The text tokenizer. c

A wrapper for Qwen3 TTS Tokenizer 25Hz/12Hz with HuggingFace-style loading. - from_pretrained(): loads speech tokenizer model via AutoModel

This is the configuration class to store the configuration of a [`Qwen3TTSTokenizerV1Config`]. It is used to instantiate a Qwen3TTSTokenizerV1Mod