web-ai-toolkit/dist/recognition-JOTKqKKk.js

AI powered features on the web made easy

import { _ as f, a as m } from "./transformers-kV3s9sef.js";
import { w as g } from "./utils-DJ16RMae.js";
let a;
function v(e, t = "Xenova/whisper-tiny") {
  return new Promise(async (s, n) => {
    try {
      a || await y(t || "Xenova/whisper-tiny", !1, "en");
      const i = new FileReader();
      i.onloadend = async () => {
        const _ = new AudioContext({
          sampleRate: 16e3
        }), p = i.result, r = await _.decodeAudioData(p);
        let l;
        if (r.numberOfChannels === 2) {
          const h = Math.sqrt(2), u = r.getChannelData(0), w = r.getChannelData(1);
          l = new Float32Array(u.length);
          for (let c = 0; c < r.length; ++c)
            l[c] = h * (u[c] + w[c]) / 2;
        } else
          l = r.getChannelData(0);
        const d = await k(l);
        s(d);
      }, i.readAsArrayBuffer(e);
    } catch (i) {
      n(i);
    }
  });
}
async function y(e = "Xenova/whisper-tiny", t, s) {
  return new Promise(async (n) => {
    a || (f.allowLocalModels = !1, f.useBrowserCache = !1, a = await m("automatic-speech-recognition", e || "Xenova/whisper-tiny", {
      // @ts-ignore
      return_timestamps: t,
      language: s,
      // @ts-ignore
      device: await g() ? "webgpu" : "wasm"
    })), n();
  });
}
async function k(e) {
  return new Promise(async (t, s) => {
    if (a) {
      const n = await a(e, {
        chunk_length_s: 30,
        stride_length_s: 5,
        // @ts-ignore
        callback_function: C,
        // after each generation step
        chunk_callback: b
        // after each chunk is processed
      });
      t(n.text);
    } else
      s();
  });
}
const o = [
  {
    tokens: [],
    finalised: !1
  }
];
function b(e) {
  const t = o[o.length - 1];
  Object.assign(t, e), t.finalised = !0, e.is_last || o.push({
    tokens: [],
    finalised: !1
  });
}
function C(e) {
  const t = a.processor.feature_extractor.config.chunk_length / a.model.config.max_source_positions, s = o[o.length - 1];
  s.tokens = [...e[0].output_token_ids];
  const n = a.tokenizer._decode_asr(o, {
    time_precision: t,
    return_timestamps: !0,
    force_full_sequences: !1
  });
  self.postMessage({
    type: "transcribe-interim",
    transcription: n[0]
  });
}
export {
  v as doLocalWhisper,
  y as loadTranscriber,
  k as localTranscribe
};