v0.3.0 — Standalone & Custom ONNX Model Support

What's New in v0.3.0

• Zero-Ceremony Functional API (webml): The default export is now a functional loader that reduces model initialization and execution to 2 lines:

  import webml from 'webml-kit';
  
  // Speech recognition (auto-detects task, normalizes mic/blob/url audio)
  const asr = await webml('onnx-community/whisper-tiny.en');
  const { text } = await asr.transcribe(audioBlob);

• Automatic Task Inference: Auto-resolves tasks (automatic-speech-recognition, text-generation, image-classification, etc.) from Hugging Face Hub metadata and model tags — zero configuration needed.
• Universal Input Coercion: Models automatically accept Blob, File, URL, AudioBuffer, and raw Float32Array without writing Web Audio API boilerplate.
• 1-Line Microphone Streaming: await asr.listen({ onTranscript: (t) => ... }) handles mic permissions, AudioContext, 16kHz resampling, and live transcription automatically.
• Standalone & Custom ONNX Support: Automatic discovery and inference for standalone ONNX models, companion audio preprocessors, and custom CTC decoders.
• WebGPU-to-WASM Fallback: Transparently recovers and executes via WASM SIMD multi-threading if WebGPU lacks operator kernels (such as 1D convolutions).