@desert-ant-labs/who

On-device speaker diarization for JavaScript that runs in the browser and in Node: who spoke when, as per-speaker turns with timestamps, in one pass over the audio. Speakers are numbered in the order they first speak, overlapping speech is marked for both, and recordings of any length are read in five-minute sections joined by voice. Audio stays local.

Two entries share one Who API:

• @desert-ant-labs/who (default): a WebAssembly pipeline with LiteRT.js inference, for the browser, on the CPU. Safe to import during server-side rendering; Who.load() runs inference only in a browser or Web Worker.
• @desert-ant-labs/who/native: a prebuilt native core, Core ML on macOS and LiteRT on Linux, for server-side inference in Node. Import it from server-only code.

npm i @desert-ant-labs/who @litertjs/core   # browser
npm i @desert-ant-labs/who                  # Node, prebuilt native core

import { Who } from "@desert-ant-labs/who";

const who = await Who.load();                   // downloads the 32 MB model on first use
const result = await who.diarize(file);         // a File or Blob, the bytes of an audio file, or samples

result.speakerCount;                            // 3
for (const turn of result.turns) {
  turn.speaker;                                 // 0, 1, 2: numbered in order of first appearance
  turn.start;                                   // 12.4, seconds from the start
  turn.end;                                     // 15.9
}
result.keepingSpeakers({ minimumShare: 0.05 }); // without the speakers under 5% of the recording
who.dispose();

diarize takes the bytes of an audio file (File, Blob, ArrayBuffer, Uint8Array), mono Float32Array samples at 16 kHz, or { samples, sampleRate } at any rate. In the browser a file is decoded with Web Audio, so any format the browser plays works. In Node the portable decoder reads WAV; decode other formats first:

import { Who } from "@desert-ant-labs/who/native";
import fs from "node:fs/promises";

const who = await Who.load();
const result = await who.diarize(await fs.readFile("call.wav"));

JSON.stringify(result) stores a diarization and Diarization.fromJSON restores it. Create one Who and reuse it; dispose() releases the model. The browser build runs the graphs on the CPU; pass accelerator: "auto" to Who.load to try WebGPU.

The full guide is in the Who docs.

License

Desert Ant Labs Source-Available License. Free for most apps, and a commercial license is required at scale. Full terms are at the link. Licensing: [email protected].

Third-party notices for the SDK: THIRD_PARTY_NOTICES.md.

Third-party notices for the weights: THIRD_PARTY_NOTICES.md.