Forward only one frame. Args: spec: shape (B, N, T=chunk_size) cache: previous chunk's last ifft frame, shape (B,

Args: audio16k: shape (b, t) audio16k_length: (b,) Returns: token: shape (b, nq, l)

Args: mask: (max_seq_len, max_seq_len) input_pos: (batch_size, seq_len) Returns: (batch_size, seq_len, max_seq

Returns: (seq_len,17), (seq_len, 17)

统计中文字符数量

统计英文单词数量

Args: tokens: (B=1, nq, L) Returns: audio: (B=1, t)

Decode codes from multiple quantizers to embeddings. Args: codes: Tensor of shape (nq, B, T) containing code indices for each

Compute the Inverse Short Time Fourier Transform (ISTFT) of a complex spectrogram. Args: spec (Tensor): Input comple

Args: tokens: (batch_size, seq_len, audio_num_codebooks+1) tokens_mask: (batch_size, seq_len, audio_num_codebooks+

Creates a triangular filter bank. Adapted from *torchaudio* and *librosa*. Args: fft_freqs (`np.ndarray` of shape `(num_f

检测文本是否包含中文字符

Args: audio: batched audio of shape (b, t)

Stream forward upsample_conv module with previous block cache. Args: x: shape (B, C, T) cache: shape (B, C, 3), w

Initialize the weights of the model. - Xavier uniform initialization for linear layers - Normal initialization for embeddings - X

Load model, add forward method, and move to device. Args: model_name_or_checkpoint_path: Name or path of pretrained model or checkpoin

Automatically discover and load node definitions

合并中文句子

合并英文句子

Setup KV caches and return a causal mask.

Returns sinusoids for positional embedding

按中文标点符号分割

按英文标点符号分割