whisper-small

Converts audio waveforms to text across 99 languages using a transformer-based encoder-decoder architecture trained on 680,000 hours of multilingual audio from the web. The model processes variable-length audio by converting it to mel-spectrograms, encoding through a 12-layer transformer encoder, and decoding via a 12-layer transformer decoder with cross-attention, outputting tokenized text that can be detokenized to readable transcriptions. Handles diverse audio conditions (background noise, accents, technical jargon) through large-scale diverse training data rather than explicit noise reduction preprocessing.

Automatically identifies the spoken language from audio input by leveraging language-specific tokens embedded in the decoder's vocabulary and learned during training on multilingual data. The model predicts a language token as the first output token after processing the audio through the encoder, enabling downstream decoding to use language-specific vocabulary and attention patterns. This detection happens implicitly during transcription without separate inference passes, making it a zero-cost auxiliary output.

Handles audio files of arbitrary length by converting them to fixed-size mel-spectrogram representations with automatic padding/truncation, enabling batch processing of heterogeneous audio lengths. The model pads shorter spectrograms to a maximum sequence length (default 3000 frames ≈ 30 seconds) and truncates longer audio, with padding tokens masked during attention computation to prevent information leakage. This design allows efficient GPU batching without reshaping individual samples.

Provides unified model weights compatible with PyTorch, TensorFlow, JAX, and ONNX runtimes through HuggingFace's transformers library abstraction layer, automatically handling framework-specific tensor operations and device placement. The model weights are stored in safetensors format (safer than pickle, faster loading) and can be loaded into any supported framework with identical numerical outputs, enabling framework-agnostic deployment and experimentation.

Supports inference in reduced-precision formats (FP16, INT8) through transformers library quantization backends, reducing model memory footprint from ~1GB (FP32) to ~500MB (FP16) or ~250MB (INT8) without retraining. The model uses post-training quantization where weights are converted to lower precision after training, with dynamic quantization of activations during inference, maintaining accuracy within 1-2% of full precision while enabling deployment on memory-constrained devices.

Processes multiple audio samples in parallel by dynamically padding each sample to the longest sequence in the batch, then using attention masks to ignore padding tokens during computation. This approach reduces wasted computation compared to padding all samples to the global maximum (3000 frames), enabling efficient batching of heterogeneous audio lengths. The implementation uses transformers' DataCollator pattern to automatically handle padding and mask generation during batch construction.

Exposes raw model logits for each predicted token, enabling downstream confidence scoring by computing softmax probabilities over the vocabulary and extracting the probability of the predicted token. This allows builders to identify low-confidence predictions, implement confidence thresholding for quality control, or generate alternative hypotheses by sampling from the probability distribution. The logits are available through the model's output structure without additional inference passes.

Enables streaming transcription by implementing sliding-window inference where overlapping audio chunks are processed sequentially with context overlap to maintain coherence across chunk boundaries. While the base model requires full audio loading, this capability describes the pattern for adapting Whisper to streaming by chunking audio into 30-second windows with 5-10 second overlap, processing each chunk independently, and merging transcriptions with overlap-based deduplication. This is not a native streaming capability but a documented inference pattern for streaming adaptation.