Decode input speech Args: methods:(List[str]): list of decoding methods to use, which could could contain the fo

Make mask tensor containing indices of padded part. See description of make_non_pad_mask. Args: lengths (torch.Tensor): Batch of len

...

log_softmax of frame activations Args: Tensor hs_pad: 3d tensor (B, Tmax, eprojs) Returns: torch.Tensor: log

fd = open_or_fd(file) Open file, gzipped file, pipe, or forward the file-descriptor. Eventually seeks in the 'file' argument contains ':offset'

argmax of frame activations Args: torch.Tensor hs_pad: 3d tensor (B, Tmax, eprojs) Returns: torch.Tensor: arg

return: wav_embedding: (b, l, v) wav_mask: (b, l), wav为有效值的位置为true

Generates responses for given prompts using Gemma model.

[key] = read_key(fd) Read the utterance-key from the opened ark/stream descriptor 'fd'.

Construct dataset from arguments We have two shuffle stage in the Dataset. The first is global shuffle at shards tar/raw file level.

Sample from logits. Args: logits: （1，1，vocab_size） temperatures: top_ps: top_ks:

Add <sos> and <eos> labels. Args: ys_pad (torch.Tensor): batch of padded target sequences (B, Lmax) sos (int): index of <sos>

封装了加载模型和分词器的逻辑 Args: checkpoint_path (str): 模型权重文件路径 config_path (str): 模型配置文件路径 device (torch.device): 加载模型的设备 R

Actually, this implements stacking frames and skipping frames. if m = 1 and n = 1, just return the origin features. if m = 1

Distort a waveform in sample point level Args: x: the origin wavefrom func: the distort function rate: sample point-level

Compute attention context vector. Args: value (torch.Tensor): Transformed value, size (#batch, ..., n_head, time2

Stable log add

Make mask tensor containing indices of non-padded part. The sequences in a batch may have different lengths. To enable batch computing, paddi

Perform padding for the list of tensors. Args: xs (List): List of Tensors [(T_1, `*`), (T_2, `*`), ..., (T_B, `*`)]. pad_value (f

Construct an PositionalEncoding object.

Transform query, key and value. Args: query (torch.Tensor): Query tensor (#batch, ..., time1, size). key (torch.Tenso

Calculate accuracy. Args: pad_outputs (Tensor): Prediction tensors (B * Lmax, D). pad_targets (LongTensor): Target label tensors

Construct an MultiHeadedAttention object.

Construct an Conv1dSubsampling object.

Apply optional mask for encoder. Args: xs (torch.Tensor): padded input, (B, L, D), L for max length mask (torch.Tensor): mask fo

Add whisper-style tokens. ([PREV] -> [previous text tokens or hotwords]).optional -- ┌-----------------------------------------------------

Returns: List[List[List[int]]]: nbest result for each utterance

Forward just one chunk Args: xs (torch.Tensor): chunk input, with shape (b=1, time, mel-dim), where `time == (ch

Args: input (torch.Tensor): [batch_size, time_step=1] padding (torch.Tensor): [batch_size,1], 1 is padding value

Generate amplitude domain mask randomly in [-100db, 0db] Args: mask_num: the slot number of the mask Returns: A list of tupl

Get a amplitude domain mask from db domain mask Args: db_mask: Optional. A list of tuple. if None, using default value. Returns:

Args: input_size (int): input dim output_size (int): dimension of attention attention_heads (int): the nu