openai-whisper

Transcribes audio in 99+ languages using a single unified encoder-decoder transformer model trained on 680,000 hours of multilingual audio from the web. The model automatically detects the spoken language without requiring explicit language specification, using a shared embedding space learned across diverse linguistic data. Inference runs locally without API calls, enabling offline transcription at scale.

Breaks audio into temporal segments and returns transcription for each segment with precise start/end timestamps and per-token confidence scores. Uses the model's internal attention mechanisms to align decoded tokens to audio frames, enabling fine-grained temporal grounding without separate alignment models. Supports both word-level and sentence-level segmentation strategies.

Transcription results can be returned as structured JSON with metadata (language, duration, segments with timestamps), enabling downstream processing without text parsing. Supports validation against JSON schemas to ensure output conforms to expected structure, useful for API contracts and data pipelines.

Provides five pre-trained model sizes (tiny, base, small, medium, large) ranging from 39MB to 3GB, enabling developers to choose optimal accuracy-speed-memory tradeoffs for their deployment constraints. Each variant uses identical architecture but different parameter counts; models are automatically downloaded and cached on first use. Supports quantization and distillation for further optimization.

Automatically handles audio format conversion, resampling, and normalization using FFmpeg as a backend. Accepts diverse input formats (MP3, WAV, M4A, FLAC, OGG, OPUS, video files) and converts to 16kHz mono PCM internally, matching the model's training data distribution. Handles variable sample rates, bit depths, and channel configurations transparently without user intervention.

Processes multiple audio files or long audio streams without loading entire files into memory simultaneously. Uses a sliding-window approach where audio is read in chunks, processed through the model, and results are yielded incrementally. Enables transcription of multi-hour audio files on systems with limited RAM by processing 30-second windows sequentially.

Supports fine-tuning pre-trained models on custom audio datasets to improve accuracy for domain-specific speech (medical terminology, accented speech, noisy environments). Uses PyTorch's standard training loop with cross-entropy loss; developers can freeze encoder layers and train only the decoder for faster convergence, or train end-to-end for maximum adaptation. Includes utilities for dataset preparation and validation.

Provides a CLI tool (`whisper` command) enabling transcription without writing Python code. Accepts audio file paths, outputs transcriptions to stdout or files, and supports flags for model selection, language specification, output format, and GPU acceleration. Useful for shell scripts, batch processing, and non-developers.

Allows specifying the language explicitly or providing a text prompt to guide decoding toward specific vocabulary and phrasing. The model uses the prompt as a conditioning signal during beam search, biasing token selection toward words and phrases appearing in the prompt. Useful for improving accuracy on domain-specific terminology or correcting common hallucinations.

Automatically detects the spoken language in audio without explicit specification, using the model's multilingual encoder to classify audio into 99+ language categories. The language detection is performed as a preliminary step before decoding, with confidence scores indicating detection certainty. Supports explicit language specification to override automatic detection if needed.

Supports GPU acceleration via PyTorch's CUDA backend and mixed-precision inference (float16) to reduce memory usage and latency. Automatically detects available GPU and uses it if present; developers can explicitly specify device placement. Mixed precision reduces model size by 50% with minimal accuracy loss, enabling larger models on memory-constrained GPUs.