Batch multiple samples for training. Processes each sample, pads sequences to same length, and organizes continuous features by modal

Frontend + Encoder. Note that this method is used by st_inference.py Args: speech: (Batch, Length, ...) speech_length

Converts an integer PCM audio signal to a floating-point representation. This function scales an integer PCM audio waveform (typically `int16

Set GPU device to use. Args: device: Device ID.

Convert special token string to multi-stream token array. Places token ID in first stream, padding tokens in other streams.

Perform Tikhonov regularization (only modifying real part). Args: mat (torch.complex64): input matrix (..., C, C) reg (float): re

Execute one training epoch.

Replace all "{shard_idx}" in a nested config with the actual shard index. This function is typically used when a config dictionary or list (e.g.,

Initialize dropout module. tie: tie dropout mask across sequence lengths (Dropout1d/2d/3d)

Construct a TooShortUttError for error handler.

Remap legacy nn.Sequential keys for subsampling modules.

Beamforming with the provided statistics. Args: data (torch.complex64/ComplexTensor): (B, F, C, T) ilens (torch.Tenso

One-step forward hypotheses. Args: hyps: Hypotheses. Returns: out: Decoder output sequences. (B, D_dec)

Create or retrieve a Dask client using the provided or global configuration. Args: config (DictConfig, optional): Cluster config. If None

Get a DataLoader for a specific range of batches. Supports endless epochs by wrapping around when batches are exhausted. If the reque

The input of normalization will be (M, C, K), where M is batch size. C is channel size and K is sequence length.

Generate a dictionary of turn-taking decisions in Human AI conversation. Args: hyp_arr (array): Lines of file containing

Generate a likelihood dictionary of turn-taking events based on predictions from the turn-taking model. Args: ref_arr (array): L

The same initialization used in DDPM.

Power based voice activity detection. Args: x: (Channel, Time) >>> x = np.random.randn(1000) >>> detect = detect_non_silence(x)

Read uttid and array in kaldi style. This function might be a bit confusing as "ark" is used for HDF5 to imitate "kaldi-rspecifier". Arg

Return the fully qualified class path for an object. Args: obj: Instantiated Python object. Returns: str: Fully qualified ``

Retrieve the IP address from an incoming HTTP request. Args: request (gr.Request): The incoming HTTP request from which

Get normalization module and arguments given parameters. Args: normalization_type: Normalization module type. eps: Value added to

Run inference. Args: text (Tensor): Input text index tensor (T_text,). feats (Tensor): Feature tensor (T_feats, aux_c

Load a user config and merge it with packaged default values. This is the higher-level helper used by recipe code. It first loads the default

Compute metrics for each test set and write a metrics JSON file. Test sets are resolved in the following order: 1. If ``metrics_config.d

Apply a function to an iterable of inputs in parallel using Dask. This helper takes care of: - Creating (or reusing) a Dask client accordin

Convert data dict to list of (role, io_type, data) messages. Either uses provided dialogue or constructs from task template. Determin

Create batches using bucket batching strategy. Samples are sorted by length and grouped into buckets such that the total tokens (max_length *

Dynamically create a parallel LLM class based on HuggingFace architecture. Creates a subclass of the original HF model with added multimodal supp

Forward one step. Args: tgt (torch.Tensor): Input token ids, int64 (#batch, maxlen_out). tgt_mask (torch.Tensor): Inp

Generate a list of long utterances from a list of short utterances.

Returns the name of this layer, e.g. 'affine1'. It does not necessarily correspond to a component name.

Get random segments. Args: x (Tensor): Input tensor (B, C, T). x_lengths (Tensor): Length tensor (B,). segment_size (int)

At the end of validation, average the top-K checkpoints and save.

Dispatch tasks to Dask and iterate over results as they complete. This helper: - Creates (or reuses) a Dask client based on the global/expl

Synthesis with PQMF. Args: x (Tensor): Input tensor (B, subbands, T // subbands). Returns: Tensor: Output te

Construct an PositionalEncoding object.

Stratified Best Fit: diversity + efficiency via interleaved packing. Items are divided into length strata, shuffled within each stratum, and

Create a minimal SpeechLMJobTemplate config dict.

Create minimal trainer_args dict.

This function replicate the entries in files like segments, utt2spk, text

Score new token batch. Args: ys (torch.Tensor): torch.int64 prefix tokens (n_batch, ylen). states (List[Any]): Scorer

Create batches using the specified batching method. Args: keys: List of sample keys to batch. key_to_length: Dictionary mapping e