Read the transcript file based on its extension. Returns: Tuple of (transcript_dict, start_timestamp, end_timestamp) Rai

Load AMI corpus audio files and transcripts. Parses the text file to extract utterance IDs and maps them to corresponding audio files

Add a new attribute to the segment. Args: key: The attribute name value: The attribute value

Get the device of the model. Returns ------- torch.device Device where the model parameters are

Load a OLMoASR model Parameters ---------- name : str one of the official model names listed in MODEL2LINK, or path

Initialize Linear layer with Kaiming normal initialization. Parameters ---------- in_features : int

Convert a timestamp in HH:MM:SS.mmm format to milliseconds. Args: timestamp: Timestamp in the format HH:MM:SS.mmm Returns: T

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

Adjust a timestamp by a specified number of milliseconds. Args: timestamp: Timestamp in HH:MM:SS.mmm format milliseconds: Number

Extract timestamp information from utterance ID. Args: utt_id: Utterance ID containing timestamp information extract_type: T

Preprocess audio waveform for validation. Resamples audio to 16kHz if needed, pads/trims to 30 seconds, and computes log mel spectrog

Preprocess text for validation. Tokenizes text using Whisper tokenizer, handles empty text with no-speech token, and creates input/ta

Convert timestamp string to milliseconds for audio processing. Converts time format used in CHiME-6 transcripts (HH:MM:SS.mmm) to

Aggregate statistics from multiple processing results. Args: statistics_list: List of statistics dictionaries Returns:

Segment audio and transcript data into chunks of up to 30 seconds. This function takes a transcript dictionary and segments it into chunks t

Extract the text from the transcript dictionary. Args: transcript: Transcript as a dictionary mapping (start, end) tuples to text

Download and prepare evaluation datasets for ASR evaluation. This function handles the automated downloading, extraction, and preprocessing o

Load and preprocess audio file with optional time-based segmentation. Loads an audio file, optionally extracts a time segment, pads/trims to

Preprocess audio arrays from HuggingFace datasets. Processes raw audio waveforms from HuggingFace dataset format, handling resampling

Save model (DDP) checkpoint Saves non-DDP and DDP model checkpoints to checkpoints/{exp_name}_{run_id} directory in the format of {file_name}_{mo

Save model (DDP) checkpoint Saves non-DDP and DDP model checkpoints to checkpoints/{exp_name}_{run_id} directory in the format of {file_name}_{mo

Add evaluation results to WandB table for detailed logging. Computes per-sample metrics and adds audio examples with predictions and ground t

Convert timestamp to token index. Args: timestamp: Timestamp in string or millisecond format timestamp_begin: Startin

Convert timestamp to token index. Args: timestamp: Timestamp in string or millisecond format timestamp_begin: Startin

Common logic for CallHome and SwitchBoard datasets. Parses HUB5 STM format files to extract audio segments with precise timing inform

Log evaluation results to multiple output destinations. Handles comprehensive logging of evaluation metrics to console, files, and WandB base

Perform subsampling on a list of lines. Args: lines: List of data lines to subsample subsample_size: Number of samples to ke

Read transcript file and extract timing and text data. Args: file_type: Type of transcript file ('vtt' or 'srt') Returns

Safely normalize text with error handling. Args: text: Text to normalize normalizer: Text normalizer Returns: N

Performs bootstrapping on the given data. Args: data (np.ndarray): The original data array. n_bootstrap (int): Number of boo

Calculate Word Error Rate (WER) and error statistics for predictions. Normalizes target and predicted text, then computes WER and detailed error

Calculate the difference between two timestamps in milliseconds. Args: timestamp1: First timestamp in HH:MM:SS.mmm format timesta

Segment transcript data without corresponding audio files. Processes transcript-only data by chunking the text content into segments. Ca

Computes mean, standard deviation, and standard error from bootstrap means. Args: bootstrap_means (np.ndarray): Array of bootstrap s

Synchronous evaluation loop for YODAS DDP training. Evaluates the model on standard datasets (LibriSpeech clean/other) using Whisper's decode

Evaluation loop with Word Error Rate (WER) calculation on multiple datasets. Evaluates the model on provided evaluation datasets using Whisper's

Extract file information from file path. Args: file_path: Path to the transcript file Returns: Dictionary containing fi

Extract base ID from utterance ID by removing segment numbers. Args: utt_id: Full utterance ID (e.g., "speaker-audio-12345-start-end

Generate predictions from model logits and decode them to text. Takes model output logits and ground truth targets, converts logits to prediction

Initialize WandB run for evaluation experiment tracking. Sets up a WandB run with appropriate configuration for ASR evaluation, inclu

Loads the WER values from a CSV file. Args: csv_file (str): Path to the CSV file. column_name (str): The name of the column

Parse subtitle content into an iterable format. Args: content: Raw subtitle content subtitle_filename: Filename to determine

Prepares the YODAS training dataset from HuggingFace datasets. Loads YODAS datasets from multiple subdirectories, concatenates them into a single

Prepares the learning rate scheduler for YODAS DDP training. Creates a LambdaLR scheduler with linear warmup and cosine decay. Also calculates

Prepares the learning rate scheduler for FSDP training. Creates a LambdaLR scheduler with linear warmup and cosine decay. Also calculates gra

Prepares the learning rate scheduler for training. Creates a LambdaLR scheduler with linear warmup and cosine decay. Also calculates gradient

Launch asynchronous evaluation subprocess for YODAS DDP training. Starts evaluation in a separate process to avoid blocking training. Only th

Launch asynchronous evaluation subprocess for FSDP training. Starts evaluation in a separate process to avoid blocking training. Only the mai

Launch asynchronous evaluation subprocess. Starts evaluation in a separate process to avoid blocking training. Only the main process (rank 0)

Save model (DDP) checkpoint Saves non-DDP and DDP model checkpoints to checkpoints/{exp_name}_{run_id} directory in the format of {file_name}_{mo

Save processing report to a log file. Args: report: Report content to save output_dir: Directory to save the log file

Initializes the distributed process group for DDP training. Sets the CUDA device for the current process and initializes the NCCL process gro

Sum multiple SegmentCounter objects efficiently. Args: counters: List of SegmentCounter objects to sum Returns: New Seg

Get the current counter value.

Build sequence with timestamps. Returns None if timestamps are invalid.

Build sequence with timestamps. Returns None if timestamps are invalid.

Classify the casing of a single caption. Args: text: Caption text to classify Returns: Casing type: 'upper', 'lower', o

Compute edit distance between texts with various fallback strategies. Args: manual_text: Original manual text machine_text:

Determine final casing type based on counts with adjustment rules. Args: casing_counts: Dictionary of casing type counts Return

Download a model from a URL and cache it locally. Parameters ---------- url : str URL to download the model from model_n

Filter machine captions to remove multi-line captions except for the last one. Args: content: Parsed WebVTT content start_in

Get the cache directory for storing downloaded models.

Get custom dataset item for long-form transcription. Loads and preprocesses audio from custom dataset formats for long-form evaluatio

Get custom dataset item for short-form transcription. Handles dataset-specific loading and preprocessing for non-HuggingFace datasets

Loads the appropriate generation config from HF repo

Get HuggingFace dataset item for long-form transcription. Processes long-form audio from HuggingFace datasets without padding or mel-

Get HuggingFace dataset item for short-form transcription. Processes audio and text from HuggingFace dataset format, handling dataset

Get item for long-form transcription. Retrieves and preprocesses audio for variable-length transcription tasks. Does not apply paddin

Extract model type and size from checkpoint path. Analyzes checkpoint file path to automatically determine model type (whisper varian

Get item for short-form transcription. Retrieves and preprocesses audio for fixed-length transcription tasks. Handles both HuggingFac

Determine the starting index for machine content processing. Args: content: Parsed WebVTT content Returns: Starting ind

Check if current text contains repeated content from history. Args: current_text: Current caption text text_history: List of

Initialize custom datasets.

Initialize HuggingFace datasets.

Load and preprocess audio file for language identification. Args: audio_file: Path to the audio file (.npy format)

Prepare content dictionary for processing. Args: line: Input line from JSONL file segment_level: Whether processing at segme

Process empty transcript case and return appropriate tokens.

Process empty transcript case and return appropriate tokens.

Process a single line through the filtering pipeline. Args: line: Data line to process config: Configuration dictionary with

Process non-empty transcript and return tokens, timestamp_mode, and updated norm_end.

Process non-empty transcript and return tokens, timestamp_mode, and updated norm_end.

Save processed lines to a compressed JSONL file. Args: output_lines: List of processed data lines original_path: Original fi

Save processing results to files. Args: output_file: Path to output file stats_file: Path to statistics file output_

Update casing statistics. Args: casing: Final casing determination content_length: Length of content stats: Statisti

Update edit distance statistics. Args: edit_distance: Calculated edit distance stats: Statistics dictionary to update

Validate sequence lengths and token indices, print warnings if issues found.

Validate sequence lengths and token indices, print warnings if issues found.

Aggregate language predictions by video ID using majority voting. Args: predicted_languages: List of (video_id, language_id) predict

Apply boolean filtering logic. Args: tag_value: The boolean value to test reference_value: The expected boolean value R

Apply categorical filtering logic. Args: tag_value: The categorical value to test reference_value: Single value or list of v