Forward pass through the complete model. Parameters ---------- mel : torch.Tensor Mel spectrogra

Main function to generate video samples from audio and transcript files. Args: wav_dir: Directory for temporary WAV files vi

Computes error bars for WER values from a CSV file. Args: csv_file (str): Path to the CSV file containing WER values. n_boot

Identify the language of manual text content. Args: content_dict: Dictionary containing manual text and length Returns:

Initialize tokenizer in each dataloader worker process. Sets up a global tokenizer instance for the worker process. This is called automatica

Initialize tokenizer in each dataloader worker process. Sets up a global tokenizer instance for the worker process. This is called automatica

Initialize tokenizer in each dataloader worker process. Sets up a global tokenizer instance for the worker process. This is called automatica

Initialize worker process with shared dataset. Args: dataset: The dataset to share across worker processes

The `MultiHeadAttention` module optionally accepts `kv_cache` which stores the key and value tensors calculated for the previous posi

Install hooks for key-value caching to enable efficient inference. The `MultiHeadAttention` module optionally accepts `kv_ca

Check if the model is multilingual. Returns ------- bool True if model supports multiple languag

Load audio files and transcripts from the dataset. Returns: Tuple[list, list]: A tuple containing: - List of audi

Load LibriSpeech audio files and transcripts. Recursively searches for transcript files and maps them to corresponding FLAC audio fil

Load Artie Bias Corpus audio files and transcripts. Parses the TSV metadata file to extract audio file paths and their corresponding

Load FLEURS dataset audio files and transcripts. Parses the test.tsv file to extract audio file paths and transcriptions for the Engl

Load VoxPopuli audio files and transcripts. Parses the ASR test TSV file to extract audio file paths and normalized transcript texts

Load CORAAL audio files and cleaned transcripts. Parses the CORAAL transcripts CSV and applies specialized text cleaning to handle di

Load CHiME-6 audio segments and transcripts. Parses JSON transcript files to extract audio segment paths and their corresponding word

Load WSJ audio extraction commands and transcripts. Parses Kaldi-format text and wav.scp files to create mappings between utterance I

Load CallHome audio segments and transcripts. Delegates to the shared HUB5 data loading logic with English prefix to extract CallHome

Load Switchboard audio segments and transcripts. Delegates to the shared HUB5 data loading logic with Switchboard prefix to extract S

Load Kincaid46 audio files and transcripts. Parses the CSV metadata file and maps entries to corresponding M4A audio files using a ze

Load CORAAL long-form audio files and transcripts. Parses the JSONL file to extract audio file paths and their corresponding full tra

Load data from a compressed JSONL file. Args: file_path: Path to the .jsonl.gz file Returns: List of JSON objects parse

Compute logits for given tokens and audio features. Parameters ---------- tokens : torch.Tensor

Evaluate ASR model performance on long-form transcription tasks. This evaluation function handles variable-length audio segments (typically longe

Main function to process YODAS dataset resegmentation. Args: input_dir: Directory containing YODAS Arrow files audio_output_

Main entry point for the data tagging pipeline. Args: config_path: Path to YAML configuration file input_dir: Directory cont

Main entry point for audio language assignment pipeline. This function orchestrates the entire process of assigning audio language tags

Main training function for YODAS DDP model training. Orchestrates the complete DDP training pipeline for YODAS including: - Distributed train

Main training function for FSDP Whisper model training with timestamps. Orchestrates the complete FSDP training pipeline including: - Distrib

Main training function for distributed Whisper model training with timestamps. Orchestrates the complete training pipeline including: - Distr

Get the number of languages supported by the model. Returns ------- int Number of supported lang

Load sample dictionaries from compressed YODAS dataset files. Currently only supports gzip (.gz) compressed JSONL files. Each line should con

Load sample dictionaries from compressed files. Supports loading JSONL data from either gzip (.gz) or zstandard (.zst) compressed files. Each

Load sample dictionaries from compressed files. Supports loading JSONL data from either gzip (.gz) or zstandard (.zst) compressed files. Each

Check if transcript text exceeds model context length. Args: timestamps: List of (start, end) timestamp tuples transcript: Transc

Wrapper function for parallel processing of audio segmentation. Unpacks arguments and calls create_segment for use with multiprocessing.

Wrapper function for parallel file list generation. Args: args: Tuple of (seg_dir, video_dir, wav_dir) Returns: List of

Wrapper function for parallel segment generation. Args: args: Tuple of (segment_info, output_dir) Returns: Generated se

Wrapper function for parallel video generation. Args: args: Tuple of (npy_file, transcript_file, output_file, wav_file) Returns

Parallel wrapper for preprocess_jsonl function. Args: args: Tuple containing all arguments for preprocess_jsonl Returns:

Wrapper function for parallel processing of transcript files. Args: file_path: Path to the transcript file Returns: Dic

Legacy method for backwards compatibility. Delegates to AudioProcessor for audio preprocessing. Maintained for compatibility with exi

Preprocess multiple JSONL files in parallel. Args: input_dir: Directory containing JSONL files to preprocess output_dir: Dir

Process audio language identification from a directory. This function loads transcript data from all .jsonl.gz files in the input directory,

Wrapper function for parallel processing of JSONL files. This function unpacks arguments and calls the main file processing function. It

Process a single JSONL file with the configured tagging pipeline. Args: jsonl_path: Path to the input JSONL.gz file config_d

Process a single JSONL file with filtering and optional subsampling. Args: jsonl_path: Path to the input JSONL.gz file outpu

Remove everything after the first instance of "<|endoftext|>" in a string. Args: text: The string to process Returns: The st

Perform reservoir sampling on a chunk of files. This function processes a subset of files and maintains a reservoir sample of the specif

Evaluate ASR model performance on short-form transcription tasks. This is the main evaluation function for fixed-length audio segments (typically

Convert transcript files to compressed JSONL format. This function processes transcript files (single file, directory, or list of files)

Validate that timestamps are within bounds and properly ordered. Args: timestamps: List of (start, end) timestamp tuples global_s

Convert the machine segment to a dictionary representation. Returns: Dictionary containing all machine segment data

Convert the counter to a dictionary representation. Returns: Dictionary containing all counter values

Check if audio array duration is too short. Args: audio_arr: Audio data as numpy array sample_rate: Audio sample rate in Hz

Check if audio duration between timestamps is too short. Args: start: Start timestamp in HH:MM:SS.mmm format end: End timestamp i

Trim an audio file to specified start and end timestamps. Trims the audio file to the specified start and end timestamps and optionally saves

Create a Video object for display in Jupyter notebooks. Args: output_file: Path to the video file Returns: IPython Vide

Write a segment of the transcript to a file. Args: audio_begin: Beginning timestamp of the audio segment timestamps: List of (sta