Process all segment directories in the data directory. Args: data_dir: Root directory containing segment subdirectories vide

Process a random sample of data from samples dictionary file. Args: samples_dicts_file: Path to JSONL file containing sample diction

Process a single transcript file and extract metadata. Args: file_path: Path to the transcript file audio_extension: Extensi

Process transcript files in parallel. Args: input_files: List of transcript file paths num_processes: Number of processes to

Process all videos in parallel and return successful results. Args: video_seg_list: List of video processing information tuples

Compute attention using manual Q, K, V computation. Parameters ---------- q : Tensor Query tenso

Read and extract a specific file from a .tar.gz archive. Args: tar_gz_path: Path to the .tar.gz archive file file_in_tar: Pa

Read an SRT file or string. Args: file_path: Path to the SRT file transcript_string: SRT transcript as a string

Read a WebVTT file. Args: file_path: Path to the WebVTT file transcript_string: WebVTT content as string Ret

Resegment dataset based on duration constraints. Groups consecutive utterances from the same speaker/source into segments that don't exc

Execute a function in parallel with progress tracking. Args: func: Function to execute in parallel argument_list: List of ar

Execute a function in parallel with progress tracking. Args: func: Function to execute in parallel argument_list: List of ar

Save percentiles to a JSON file. Args: percentiles: List of percentile values to save output_path: Path where to save the pe

Save processed JSONL data to a compressed file. Args: data: List of JSON objects to save output_path: Path where to save the

Convert parsed subtitle content back to string format. Args: subtitle_content: Parsed subtitle object (WebVTT or SRT) Returns:

Initializes the distributed process group for FSDP training. Sets the CUDA device for the current process and initializes the NCCL process gr

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

Sets up Weights and Biases logging for YODAS DDP training. Initializes W&B with comprehensive training configuration and defines custom metri

Sets up Weights and Biases logging for FSDP training. Initializes W&B with comprehensive training configuration and defines custom metrics fo

Sets up the Weights and Biases logging Args: run_id: The run ID exp_name: The experiment name job_type: The type of job

This function is from OpenAI's Whisper repository. The original version can be found at: https://github.com/openai/whisper/blob/ba3f3cd54b0e5

This function is from OpenAI's Whisper repository. The original version can be found at: https://github.com/openai/whisper/blob/ba3f3cd54b0e5

Main DDP training loop for YODAS with automatic mixed precision. Performs forward pass, backward pass, gradient accumulation, and optional evalua

Main FSDP training loop with automatic mixed precision and distributed training. Performs forward pass, backward pass, gradient accumulation, and

Main training loop with automatic mixed precision and distributed training. Performs forward pass, backward pass, gradient accumulation, and opti

Transcribe an audio file using Whisper Parameters ---------- model: Whisper The Whisper model instance audio: Union[str

Unarchive a .jsonl.gz file and optionally save the uncompressed content. Args: file_path: Path to the .jsonl.gz file output_

Run validation on multiple HuggingFace speech datasets. Evaluates the FSDP model on validation sets including TED-LIUM, VoxPopuli, Common Voi

Write transcript dictionaries to compressed JSONL file. Args: transcript_dicts: List of transcript metadata dictionaries out

Get a single sample from the dataset. Args: idx: Sample index Returns: Tuple of (video_id, audio_ar

Get a single sample from the dataset. Dispatches to appropriate method based on the transcription task type. Returns preprocessed aud

Get a single sample from the YODAS dataset. Processes both audio and text for YODAS training, including timestamp processing and rand

Get a single sample from the dataset. Returns: Tuple containing: - audio_file: path to audio file - trans

Get a single validation sample. Extracts audio and text from the dataset, handles dataset-specific field names, and preprocesses both

Get a single sample from the dataset. Returns: Tuple containing: - audio_file: path to audio file - trans

Initialize the shared counter. Args: initial_value: Starting value for the counter (default: 0)

Initialize the dataset. Args: data: List of transcript segment dictionaries containing audio file paths

Initialize the dataset loader with root directory. Args: root_dir (str): Path to the root directory containing dataset files

Initialize the evaluation dataset. Sets up the appropriate dataset loader based on the evaluation set and task type. Handles both Hug

Initialize the YODAS dataset. Args: hf_dataset: HuggingFace dataset with YODAS format samples n_text_ctx: Maximum tex

Initialize validation dataset. Args: n_text_ctx: Maximum number of text context tokens val_set: Name of the HuggingFa

Initialize TranscriptReader with either file path or transcript string. Args: file_path: Path to transcript file tran

Initialize a Segment object. Args: subtitle_file: Path to the subtitle file seg_content: Content of the segm

Initialize a MachineSegment object. Args: subtitle_file: Path to the subtitle file seg_content: Content of t

Initialize a SegmentCounter object. Args: segment_count: Total number of valid segments processed over_30_li

Initialize 1D convolution layer with Kaiming normal initialization. Parameters ---------- in_channels : int

Initialize multi-head attention layer. Parameters ---------- n_state : int Dimension of the mode

Initialize residual attention block. Parameters ---------- n_state : int Dimension of the model

Initialize audio encoder. Parameters ---------- n_mels : int Number of mel-frequency channels

Initialize text decoder. Parameters ---------- n_vocab : int Vocabulary size n_ctx : int

Initialize OLMoASR model. Parameters ---------- dims : ModelDimensions Model dimensions configur

Return the number of samples in the dataset.

Return the number of samples in the dataset. Returns: int: Total number of audio samples available for evaluation Note:

Return the number of samples in the dataset.

Return the number of samples in the validation dataset.

Forward convolution with dtype casting. Parameters ---------- x : Tensor Input tensor we

Format timestamp components into HH:MM:SS.mmm format.

Analyze the casing pattern of subtitle content. Args: content_dict: Dictionary containing content iterator and length Returns:

Add language tag to a sample dictionary. Args: sample_dict: Original sample dictionary language_id: Identified language code

Calculate edit distance between manual and machine text. Args: content_dict: Dictionary containing manual and machine text n

Calculate the Word Error Rate (WER) between two strings. Args: pair: Tuple of (reference_text, hypothesis_text) Returns: Wor

Check if content contains both commas and periods. Args: content_dict: Dictionary containing content iterator and length Return

Check if content has proper capitalization after punctuation. Args: content_dict: Dictionary containing content iterator and length

Segment audio and transcript files from Google Cloud Storage tar.gz archives. Processes files stored in tar.gz archives on GCS by extracting

Segment local audio and transcript files into chunks. Processes local files by reading the transcript and segmenting both audio and tran

Convert a single timestamp to token index.

Count the total number of words in content. Args: content_dict: Dictionary containing content iterator Returns: Tuple o

Detect if content has repeating lines. Args: content_dict: Dictionary containing content iterator and length Returns: T

Embed audio features using the encoder. Parameters ---------- mel : torch.Tensor Mel spectrogram

x : torch.Tensor, shape = (batch_size, n_mels, n_ctx) the mel spectrogram of the audio

x : torch.LongTensor, shape = (batch_size, <= n_ctx) the text tokens xa : torch.Tensor, shape = (batch_size, n_audio_ctx,

Forward pass through layer normalization. Parameters ---------- x : Tensor Input tensor to norma

Forward pass through multi-head attention. Parameters ---------- x : Tensor Input tensor for que

Forward pass through residual attention block. Parameters ---------- x : Tensor Input tensor

Forward pass through audio encoder. Parameters ---------- x : torch.Tensor, shape = (batch_size, n_mels, n_c

Forward pass through text decoder. Parameters ---------- x : torch.LongTensor, shape = (batch_size, <= n_ctx