distil-large-v3

Converts audio streams into text across 99 languages using a distilled Whisper encoder-decoder architecture that reduces the original Whisper model by ~49% while maintaining accuracy. The model uses cross-attention between audio mel-spectrogram features and learned token embeddings, processing variable-length audio through a convolutional feature extractor followed by transformer layers. Distillation was applied via knowledge transfer from the full Whisper large model, enabling efficient inference on CPU and edge devices.

Automatically detects the spoken language in audio input by analyzing the acoustic features through the encoder portion of the distilled Whisper model, which learns language-specific phonetic patterns during training. The model outputs language probabilities across 99 supported languages, allowing downstream systems to route transcription or handle multilingual content appropriately. Language detection occurs as a byproduct of the transcription process without additional inference passes.

Enables efficient inference on CPU and edge devices through support for multiple model formats (PyTorch, JAX, ONNX) and quantization strategies. The model can be loaded in float32, float16, or quantized int8 formats depending on hardware constraints, with ONNX export enabling runtime optimization via ONNX Runtime's graph optimization and operator fusion. The distilled architecture (49% smaller than Whisper large) combined with quantization can reduce memory footprint to <1GB, enabling deployment on devices with limited RAM.

Processes multiple audio files of varying lengths in a single inference pass by padding shorter sequences and masking padded positions in the attention mechanism. The model's convolutional feature extractor handles variable-length mel-spectrograms, and the transformer encoder uses attention masks to prevent the model from attending to padding tokens. Batch processing reduces per-sample overhead and enables efficient GPU/CPU utilization when processing datasets.

Exports the distilled Whisper model to ONNX (Open Neural Network Exchange) format, enabling inference across diverse platforms (Windows, Linux, macOS, mobile, web browsers) using ONNX Runtime. The export process converts PyTorch operations to ONNX opset 14+, preserving the encoder-decoder architecture and attention mechanisms. ONNX Runtime applies graph-level optimizations (operator fusion, constant folding) and supports hardware-specific execution providers (CPU, GPU, CoreML for iOS, NNAPI for Android).

Extracts precise timing information for each generated token (word or subword) by tracking the decoder's output positions and mapping them back to input audio timestamps. The model outputs token-level alignments through the decoder's attention weights over the encoder output, enabling applications to determine exactly when each word was spoken. This is achieved by preserving the encoder-decoder attention patterns during inference and post-processing them to align tokens with audio frames.