webmos

Run DNSMOS (Mean Opinion Score) entirely in the browser using ONNX Runtime Web.

No server, no setup, no file copying — just import and go. Model and WASM files load from CDN automatically.

See it in action: webmos-demo

Install

npm install webmos

Usage

import { runDNSMOS } from "webmos";

const result = await runDNSMOS(audioFloat32Array, sampleRate);

console.log(result);
// {
//   mos_sig:  3.82,   — Signal quality (1–5)
//   mos_bak:  4.12,   — Background noise quality (1–5)
//   mos_ovr:  3.42,   — Overall quality (1–5)
// }

That's it. The model loads automatically on the first call.

Pre-warming (optional)

If you want to load the model ahead of time (e.g. while the user is selecting a file):

import { initDNSMOS } from "webmos";

await initDNSMOS(); // pre-loads model from CDN

API

runDNSMOS(audioData: Float32Array, sampleRate: number): Promise<DNSMOSResult>

Run DNSMOS inference on raw PCM audio. Auto-initialises on first call.

• audioData — Mono float32 samples (any sample rate; resampled to 16 kHz internally)
• sampleRate — Sample rate of the input audio

Returns:

interface DNSMOSResult {
  mos_sig: number;  // Signal quality (1–5)
  mos_bak: number;  // Background noise quality (1–5)
  mos_ovr: number;  // Overall quality (1–5)
}

initDNSMOS(options?: DNSMOSOptions): Promise<void>

Optional. Pre-load the model. Called automatically by runDNSMOS() if not called explicitly.

| Option | Type | Default | Description | |--------|------|---------|-------------| | modelUrl | string | jsdelivr CDN | URL to sig_bak_ovr.onnx | | wasmPaths | string | jsdelivr CDN | ONNX Runtime WASM file URL prefix | | ortLogLevel | string | "error" | Log level: "verbose", "info", "warning", "error", "fatal" |

How It Works

1. Audio is resampled to 16 kHz if needed
2. Raw audio is fed directly to sig_bak_ovr.onnx running in WebAssembly
3. Model outputs are calibrated via polynomial fitting to produce MOS scores

No DSP preprocessing — the model takes raw audio directly.

License

ISC