Whisper CLI

Transcribes audio in 98 languages to text in the original language using a unified Transformer sequence-to-sequence architecture with a shared AudioEncoder that processes mel spectrograms into language-agnostic embeddings, then a TextDecoder that generates tokens autoregressively. The system handles variable-length audio by padding or trimming to 30-second segments and uses task-specific tokens to signal transcription mode, enabling a single model to handle multiple languages without language-specific branches.

Translates non-English speech directly to English text by using a task-specific token in the TextDecoder that signals translation mode, bypassing the need for intermediate transcription-then-translation pipelines. The AudioEncoder processes mel spectrograms identically to transcription, but the decoder generates English tokens directly from audio embeddings, reducing latency and error propagation compared to cascaded systems.

Exposes two levels of API abstraction: a high-level transcribe() function that handles end-to-end transcription with automatic audio loading, preprocessing, and result formatting, and a low-level decode() function that provides fine-grained control over decoding options (beam width, temperature, language constraints). The high-level API is suitable for simple use cases, while the low-level API enables advanced customization for researchers and developers building complex pipelines.

Detects the spoken language in audio by generating a language token from the AudioEncoder embeddings before decoding text, using the model's multilingual training to recognize acoustic patterns distinctive to each language. The system identifies language during the initial decoding step and can be queried directly via the language identification task token, enabling language detection without full transcription.

Converts raw audio files in multiple formats (MP3, WAV, M4A, FLAC, OGG) to mel-spectrogram features via FFmpeg decoding and log-scale mel-frequency filtering, then normalizes variable-length audio to fixed 30-second segments via padding or trimming. The pipeline uses whisper.load_audio() for format-agnostic decoding, whisper.pad_or_trim() for segment normalization, and whisper.log_mel_spectrogram() for feature extraction, enabling the model to process diverse audio sources with consistent preprocessing.

Generates transcription or translation tokens autoregressively using a TextDecoder that processes AudioEncoder embeddings and previously generated tokens, with support for multiple decoding strategies including greedy decoding, beam search, and temperature-based sampling. The system uses a sliding-window context approach to handle audio longer than 30 seconds by processing overlapping segments and merging results, and supports DecodingOptions for fine-grained control over decoding behavior (beam width, temperature, language constraints).

Generates precise word-level timestamps by aligning decoded tokens to audio segments using the model's internal attention weights and token probabilities, enabling subtitle generation and fine-grained audio-text synchronization. The system decodes text at the segment level (30 seconds), then uses token timing information to map each word back to its position in the original audio, producing timestamps accurate to ~100ms granularity.

Provides six model sizes (tiny, base, small, medium, large, turbo) with parameter counts ranging from 39M to 1550M, enabling users to select optimal speed-accuracy tradeoffs based on hardware constraints and latency requirements. Each model has English-only variants (tiny.en, base.en, small.en) that sacrifice multilingual capability for 10-40% speed improvement, and the turbo model (809M) optimizes large-v3 for 8x faster inference with minimal accuracy degradation but no translation support.

Accelerates inference on NVIDIA GPUs by leveraging PyTorch's CUDA backend for AudioEncoder and TextDecoder operations, with optional mixed-precision (float16) support to reduce memory usage and increase throughput. The system automatically detects CUDA availability and moves model weights to GPU if available, and supports batch processing of multiple audio files on GPU for higher throughput than sequential CPU inference.

Provides a command-line interface (whisper command) that wraps the Python API with support for multiple output formats (TXT, JSON, VTT, SRT), batch processing of multiple audio files, and configurable transcription options (model size, language, task type). The CLI uses argparse for argument parsing and supports both single-file and directory-based batch processing, with output files automatically named based on input filenames.

Controls transcription behavior (language, task type) by prepending special tokens to the TextDecoder input, allowing the same model weights to handle different languages and tasks (transcription vs. translation) without separate model branches. The system uses reserved token IDs to signal language (e.g., <|en|> for English) and task (e.g., <|transcribe|> vs. <|translate|>), enabling fine-grained control over model behavior at inference time.