Setup forward pass to return attention weights from alignment heads as output

Transcribe an audio file using PyTorch OpenAI Whisper

Decomposition of `torch.stft` using non-complex dtypes Reference: https://github.com/pseeth/torch-stft

Given a `WhisperTextDecoder` instance, pre-compute the autoregressive decoder loop using forced decoder context prefix tokens and save results

If size of file at `audio_file_path` is larger than OpenAI API max file size, compress with ffmpeg

Prepare random test inputs with: - Key-value cache - Fixed length encoder context utilization - Randomized length decoder context uti

f""" Evaluates models from {MODEL_REPO_ID} on benchmark datasets and publishes results to hf.co/datasets/{EVALS_REPO_ID}

Evaluate the given whisper pipeline implementation on a registered dataset.

Evaluate a single audio file with whisper_pipeline

AI Hub compile options recommended for the model

Load a Whisper model from either Hugging Face hub or local path Args: model_path: Either a Hugging Face model ID or local directory path

Same function as `librosa.util.pad_center`

Move quantized variants from nested folders into publishable structure

Replace any other markers, symbols, and punctuations with a space, and drop any diacritics (category 'Mn' and some manual mappings)

Args: task: 0->transcribe, 1->translate language: Original token index from vocab, e.g. <|en|>=50259

Args: task_idx: Refers to the flattened index representing selected task & language

Transcribe an audio file using the Whisper pipeline

Transcribe an audio folder using the WhisperKit CLI

Enumerates all TextDecoder context prefixes that define a valid task, precomputes the KV cache values and initializes an `nn.Embeddin

Helper function to run the HuggingFace WhisperDecoder implementation in prefill mode using the I/O format of Argmax WhisperDecoder implementa

Build and return inputs for the first decoding step with empty caches. Should be finalized with <|startoftranscript|> query tokens downstr

Run the decoder on current inputs and update kv caches