whisper-base

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 uses mel-spectrogram feature extraction on the audio input, processes it through a 12-layer transformer encoder, and generates text tokens via a 12-layer transformer decoder with cross-attention, enabling robust transcription without language-specific fine-tuning.

Identifies the spoken language in audio by processing mel-spectrograms through the transformer encoder and classifying the resulting embeddings against 99 language tokens without explicit language labels. The model learns language-specific acoustic patterns during training on multilingual web audio, enabling implicit language detection as a byproduct of the transcription task.

Automatically handles diverse audio formats and sample rates by converting input audio to 16kHz mono waveforms and computing mel-spectrograms (80 mel-frequency bins, 400ms window, 160ms stride) as fixed-size feature representations. The preprocessing pipeline uses librosa's resampling and mel-scale filterbank computation, normalizing audio to a standard format that the transformer encoder expects, with automatic gain control via log-amplitude scaling.

Processes multiple audio files of different lengths in a single batch by padding shorter sequences to match the longest sequence in the batch, computing mel-spectrograms for all audios, and running the transformer encoder-decoder in parallel. The implementation uses attention masks to ignore padded positions, enabling efficient GPU utilization while handling variable-length inputs without truncation or resampling.

Provides unified model weights and inference APIs compatible with PyTorch, TensorFlow, and JAX through HuggingFace's transformers library abstraction layer. The model is distributed in SafeTensors format (a safe, fast serialization standard) with framework-specific weight loading, allowing developers to choose their preferred framework without retraining or format conversion.

Supports inference on resource-constrained devices (mobile, edge) through quantization to 8-bit or 16-bit precision using PyTorch's quantization APIs or ONNX Runtime quantization. Quantized models reduce memory footprint from 300MB (float32) to ~75MB (int8) and accelerate inference by 2-4x on CPU, enabling deployment on devices with <1GB RAM.