@runanywhere/web-onnx

The independent ONNX Runtime and Sherpa-ONNX backend for the RunAnywhere Web SDK. It provides embeddings, Speech-to-Text, Text-to-Speech, and Voice Activity Detection through the public RunAnywhere facade.

This browser package integrates with core only through @runanywhere/web/backend; it does not depend on the llama.cpp backend or core's private /internal entrypoint.

Install and initialize

npm install @runanywhere/web @runanywhere/web-onnx

import { RunAnywhere, SDKEnvironment } from '@runanywhere/web';
import { ONNX } from '@runanywhere/web-onnx';

await RunAnywhere.initialize({
  environment: SDKEnvironment.SDK_ENVIRONMENT_DEVELOPMENT,
});
await ONNX.register();

// Register/download/load compatible models through RunAnywhere first.
const transcript = await RunAnywhere.transcribe(audioSamples, {
  sampleRate: 16_000,
});
await RunAnywhere.speak('Speech synthesized locally in the browser.');

Capabilities

• Whisper and other Sherpa-compatible STT models, including batch and streamed partial/final results.
• Piper/Sherpa-compatible TTS synthesis and browser playback.
• Silero/Sherpa-compatible VAD.
• ONNX embeddings used by the cross-backend Web RAG provider.

Artifact

The package ships one self-contained pair:

wasm/racommons-onnx-sherpa.{js,wasm}

It includes the required ONNX and Sherpa registration exports and owns its native module lifetime. It does not reuse the core or llama.cpp WASM module. Configure your bundler to retain package-relative import.meta.url resolution. Vite users should exclude @runanywhere/web-onnx from dependency pre-bundling.

Build and verify from sdk/runanywhere-web:

npm run build:wasm -- --onnx
npm run build -w packages/onnx
npm run verify:package -w packages/onnx

This is a browser-bundler package, not a server-side Node.js inference runtime.

License

Copyright (c) 2025 RunAnywhere, Inc. See the included LICENSE file for the custom RunAnywhere License.