mellon 0.0.18 → 0.0.20

package/README.md

@@ -17,7 +17,7 @@ Offline, fully in-browser **hotword / wake-word detection** powered by [Efficien
 2. [Quick start](#quick-start)
 3. [Enrolling words](#enrolling-custom-words)
 4. [API reference](#api-reference)
-   - [Mellon](#mellon)
+   - [Detector](#detector)
    - [EnrollmentSession](#enrollmentsession)
 5. [Science behind the lib](#science-behind-the-lib)
 ---
@@ -31,9 +31,9 @@ npm install mellon
 ## Quick start
 
 ```js
-import { Mellon } from 'mellon'
+import { Detector } from 'mellon'
 
-const hotWordDetection = new Mellon([
+const hotWordDetection = new Detector([
   {
     name: 'openDoors',
     triggers: [{ name: 'mellon', defaultRefPath: '/mellon-assets/mellon_ref.json' }],
@@ -71,9 +71,9 @@ await hotWordDetection.start() // opens the mic and listens for all registered t
 ## Enrolling custom words
 
 ```js
-import { Mellon, EnrollmentSession } from 'mellon'
+import { Detector, EnrollmentSession, Storage } from 'mellon'
 
-const hotwordDetection = new Mellon([{
+const hotwordDetection = new Detector([{
   name: 'startEngine',
   triggers: [{ name: 'start' }],
   onMatch: (triggerNameMatched, confidence) => { console.log('starting engine...') }
@@ -99,19 +99,19 @@ hotwordDetection.addCustomWord(ref)
 await hotwordDetection.start()
 
 // 4b. Persist for future sessions
-Mellon.saveWord(ref)
+Storage.saveWord(ref)
 ```
 
 ---
 
 ## API reference
 
-### `Mellon` 
+### `Detector` 
 
 The easiest way to use the library. Wraps mic access, AudioWorklet wiring, and detector management into a single class.
 
 ```ts
-class Mellon {
+class Detector {
   constructor(commands: Command[], config?: MellonConfig)
   readonly threshold:  number   // read/write; persisted in localStorage
   readonly listening:  boolean
@@ -121,7 +121,19 @@ class Mellon {
   stop(): Promise<void>
   addCustomWord(ref: WordRef): void
 
-  // Storage helpers — static, work without a Mellon instance
+  // Storage helpers — static, work without a Detector instance
+  static loadWords(storageKey?: string): WordRef[]
+  static saveWord(ref: WordRef, storageKey?: string): void
+  static deleteWord(wordName: string, storageKey?: string): void
+}
+```
+
+### `Storage`
+
+Static helpers for persisting enrolled word references in `localStorage`.
+
+```ts
+class Storage {
   static loadWords(storageKey?: string): WordRef[]
   static saveWord(ref: WordRef, storageKey?: string): void
   static deleteWord(wordName: string, storageKey?: string): void

package/dist/index.d.ts

@@ -1,5 +1,7 @@
-import { Mellon } from './Mellon';
+import { Detector } from './Detector';
 import { EnrollmentSession } from './EnrollmentSession';
+import { Storage } from './Storage';
+import { AudioUtils } from './AudioUtils';
 export type TriggerName = string;
 export interface Trigger {
     name: TriggerName;
@@ -10,28 +12,37 @@ export interface Command {
     triggers: Trigger[];
     onMatch?: (trigger: TriggerName, confidence: number) => any;
 }
-export interface MellonConfig {
+export interface DetectorConfig {
     refsStorageKey?: string;
     thresholdStorageKey?: string;
     wasmPaths?: string;
     modelPath?: string;
     audioProcessorPath?: string;
     ortCdnUrl?: string;
+    audioUtils?: AudioUtils;
 }
 export interface EnrollmentSessionConfig {
     wasmPaths?: string;
     modelPath?: string;
     ortCdnUrl?: string;
+    audioUtils?: AudioUtils;
 }
 export interface WordRef {
     word_name: TriggerName;
     model_type?: string;
     embeddings: number[][];
 }
+/**
+ * Called during {@link Detector.init} to report real download progress.
+ * @param downloaded - total bytes received so far across all assets
+ * @param total      - sum of known Content-Length values for all assets;
+ *                     may still be 0 early on (before first header is received)
+ */
+export type ProgressCallback = (downloaded: number, total: number) => void;
 export declare const DEFAULT_WASM_PATHS = "https://cdn.jsdelivr.net/npm/onnxruntime-web@1.24.3/dist/";
 export declare const DEFAULT_ORT_CDN_URL = "https://cdn.jsdelivr.net/npm/onnxruntime-web@1.24.3/dist/ort.wasm.min.mjs";
 export declare const DEFAULT_MODEL_PATH = "https://huggingface.co/ComicScrip/mellon/resolve/main/model.onnx";
 export declare const DEFAULT_AUDIO_PROCESSOR_PATH = "https://cdn.jsdelivr.net/npm/mellon@0.0.14/dist/assets/audio-processor.js";
 export declare const DEFAULT_REFS_STORAGE_KEY = "mellon-refs";
 export declare const DEFAULT_THRESHOLD_STORAGE_KEY = "mellon-threshold";
-export { Mellon, EnrollmentSession };
+export { Detector, EnrollmentSession, Storage, AudioUtils };

package/dist/mellon.cjs

@@ -1 +1 @@
-"use strict";Object.defineProperty(exports,Symbol.toStringTag,{value:"Module"});function Dt(m){return m&&m.__esModule&&Object.prototype.hasOwnProperty.call(m,"default")?m.default:m}var lt,ut;function Ut(){if(ut)return lt;ut=1;function m(o){if(this.size=o|0,this.size<=1||(this.size&this.size-1)!==0)throw new Error("FFT size must be a power of two and bigger than 1");this._csize=o<<1;for(var t=new Array(this.size*2),s=0;s<t.length;s+=2){const i=Math.PI*s/this.size;t[s]=Math.cos(i),t[s+1]=-Math.sin(i)}this.table=t;for(var n=0,r=1;this.size>r;r<<=1)n++;this._width=n%2===0?n-1:n,this._bitrev=new Array(1<<this._width);for(var a=0;a<this._bitrev.length;a++){this._bitrev[a]=0;for(var e=0;e<this._width;e+=2){var h=this._width-e-2;this._bitrev[a]|=(a>>>e&3)<<h}}this._out=null,this._data=null,this._inv=0}return lt=m,m.prototype.fromComplexArray=function(t,s){for(var n=s||new Array(t.length>>>1),r=0;r<t.length;r+=2)n[r>>>1]=t[r];return n},m.prototype.createComplexArray=function(){const t=new Array(this._csize);for(var s=0;s<t.length;s++)t[s]=0;return t},m.prototype.toComplexArray=function(t,s){for(var n=s||this.createComplexArray(),r=0;r<n.length;r+=2)n[r]=t[r>>>1],n[r+1]=0;return n},m.prototype.completeSpectrum=function(t){for(var s=this._csize,n=s>>>1,r=2;r<n;r+=2)t[s-r]=t[r],t[s-r+1]=-t[r+1]},m.prototype.transform=function(t,s){if(t===s)throw new Error("Input and output buffers must be different");this._out=t,this._data=s,this._inv=0,this._transform4(),this._out=null,this._data=null},m.prototype.realTransform=function(t,s){if(t===s)throw new Error("Input and output buffers must be different");this._out=t,this._data=s,this._inv=0,this._realTransform4(),this._out=null,this._data=null},m.prototype.inverseTransform=function(t,s){if(t===s)throw new Error("Input and output buffers must be different");this._out=t,this._data=s,this._inv=1,this._transform4();for(var n=0;n<t.length;n++)t[n]/=this.size;this._out=null,this._data=null},m.prototype._transform4=function(){var t=this._out,s=this._csize,n=this._width,r=1<<n,a=s/r<<1,e,h,i=this._bitrev;if(a===4)for(e=0,h=0;e<s;e+=a,h++){const v=i[h];this._singleTransform2(e,v,r)}else for(e=0,h=0;e<s;e+=a,h++){const v=i[h];this._singleTransform4(e,v,r)}var l=this._inv?-1:1,c=this.table;for(r>>=2;r>=2;r>>=2){a=s/r<<1;var d=a>>>2;for(e=0;e<s;e+=a)for(var _=e+d,u=e,f=0;u<_;u+=2,f+=r){const v=u,p=v+d,g=p+d,w=g+d,T=t[v],F=t[v+1],A=t[p],y=t[p+1],b=t[g],S=t[g+1],P=t[w],C=t[w+1],E=T,R=F,M=c[f],D=l*c[f+1],U=A*M-y*D,z=A*D+y*M,j=c[2*f],L=l*c[2*f+1],B=b*j-S*L,K=b*L+S*j,N=c[3*f],O=l*c[3*f+1],k=P*N-C*O,G=P*O+C*N,$=E+B,x=R+K,I=E-B,Y=R-K,J=U+k,H=z+G,W=l*(U-k),Q=l*(z-G),tt=$+J,rt=x+H,et=$-J,ot=x-H,nt=I+Q,at=Y-W,it=I-Q,ct=Y+W;t[v]=tt,t[v+1]=rt,t[p]=nt,t[p+1]=at,t[g]=et,t[g+1]=ot,t[w]=it,t[w+1]=ct}}},m.prototype._singleTransform2=function(t,s,n){const r=this._out,a=this._data,e=a[s],h=a[s+1],i=a[s+n],l=a[s+n+1],c=e+i,d=h+l,_=e-i,u=h-l;r[t]=c,r[t+1]=d,r[t+2]=_,r[t+3]=u},m.prototype._singleTransform4=function(t,s,n){const r=this._out,a=this._data,e=this._inv?-1:1,h=n*2,i=n*3,l=a[s],c=a[s+1],d=a[s+n],_=a[s+n+1],u=a[s+h],f=a[s+h+1],v=a[s+i],p=a[s+i+1],g=l+u,w=c+f,T=l-u,F=c-f,A=d+v,y=_+p,b=e*(d-v),S=e*(_-p),P=g+A,C=w+y,E=T+S,R=F-b,M=g-A,D=w-y,U=T-S,z=F+b;r[t]=P,r[t+1]=C,r[t+2]=E,r[t+3]=R,r[t+4]=M,r[t+5]=D,r[t+6]=U,r[t+7]=z},m.prototype._realTransform4=function(){var t=this._out,s=this._csize,n=this._width,r=1<<n,a=s/r<<1,e,h,i=this._bitrev;if(a===4)for(e=0,h=0;e<s;e+=a,h++){const ht=i[h];this._singleRealTransform2(e,ht>>>1,r>>>1)}else for(e=0,h=0;e<s;e+=a,h++){const ht=i[h];this._singleRealTransform4(e,ht>>>1,r>>>1)}var l=this._inv?-1:1,c=this.table;for(r>>=2;r>=2;r>>=2){a=s/r<<1;var d=a>>>1,_=d>>>1,u=_>>>1;for(e=0;e<s;e+=a)for(var f=0,v=0;f<=u;f+=2,v+=r){var p=e+f,g=p+_,w=g+_,T=w+_,F=t[p],A=t[p+1],y=t[g],b=t[g+1],S=t[w],P=t[w+1],C=t[T],E=t[T+1],R=F,M=A,D=c[v],U=l*c[v+1],z=y*D-b*U,j=y*U+b*D,L=c[2*v],B=l*c[2*v+1],K=S*L-P*B,N=S*B+P*L,O=c[3*v],k=l*c[3*v+1],G=C*O-E*k,$=C*k+E*O,x=R+K,I=M+N,Y=R-K,J=M-N,H=z+G,W=j+$,Q=l*(z-G),tt=l*(j-$),rt=x+H,et=I+W,ot=Y+tt,nt=J-Q;if(t[p]=rt,t[p+1]=et,t[g]=ot,t[g+1]=nt,f===0){var at=x-H,it=I-W;t[w]=at,t[w+1]=it;continue}if(f!==u){var ct=Y,yt=-J,Tt=x,At=-I,Ft=-l*tt,bt=-l*Q,St=-l*W,Pt=-l*H,Ct=ct+Ft,Et=yt+bt,Rt=Tt+Pt,Mt=At-St,ft=e+_-f,vt=e+d-f;t[ft]=Ct,t[ft+1]=Et,t[vt]=Rt,t[vt+1]=Mt}}}},m.prototype._singleRealTransform2=function(t,s,n){const r=this._out,a=this._data,e=a[s],h=a[s+n],i=e+h,l=e-h;r[t]=i,r[t+1]=0,r[t+2]=l,r[t+3]=0},m.prototype._singleRealTransform4=function(t,s,n){const r=this._out,a=this._data,e=this._inv?-1:1,h=n*2,i=n*3,l=a[s],c=a[s+n],d=a[s+h],_=a[s+i],u=l+d,f=l-d,v=c+_,p=e*(c-_),g=u+v,w=f,T=-p,F=u-v,A=f,y=p;r[t]=g,r[t+1]=0,r[t+2]=w,r[t+3]=T,r[t+4]=F,r[t+5]=0,r[t+6]=A,r[t+7]=y},lt}var zt=Ut();const xt=Dt(zt);class pt{constructor(o=16e3,t=512,s=64){this.sampleRate=o,this.nfft=t,this.nfilt=s,this.fft=new xt(t),this.melFilters=this.createMelFilterbank()}hzToMel(o){return 2595*Math.log10(1+o/700)}melToHz(o){return 700*(10**(o/2595)-1)}createMelFilterbank(){const t=this.sampleRate/2,s=this.hzToMel(0),n=this.hzToMel(t),r=new Float32Array(this.nfilt+2);for(let i=0;i<this.nfilt+2;i++)r[i]=s+i*(n-s)/(this.nfilt+1);const e=r.map(i=>this.melToHz(i)).map(i=>Math.floor((this.nfft+1)*i/this.sampleRate)),h=[];for(let i=0;i<this.nfilt;i++){const l=new Float32Array(Math.floor(this.nfft/2)+1);for(let c=e[i];c<e[i+1];c++)l[c]=(c-e[i])/(e[i+1]-e[i]);for(let c=e[i+1];c<e[i+2];c++)l[c]=(e[i+2]-c)/(e[i+2]-e[i+1]);h.push(l)}return h}logfbank(o){const t=Math.floor(.025*this.sampleRate),s=Math.floor(.01*this.sampleRate),n=1+Math.ceil((o.length-t)/s),r=new Float32Array(n*this.nfilt),a=new Float32Array(this.nfft),e=this.fft.createComplexArray();for(let h=0;h<n;h++){const i=h*s;a.fill(0);for(let d=0;d<t&&i+d<o.length;d++)a[d]=o[i+d];const l=this.fft.toComplexArray(a,null);this.fft.transform(e,l);const c=new Float32Array(Math.floor(this.nfft/2)+1);for(let d=0;d<c.length;d++){const _=e[2*d],u=e[2*d+1];c[d]=1/this.nfft*(_*_+u*u),c[d]===0&&(c[d]=1e-30)}for(let d=0;d<this.nfilt;d++){let _=0;const u=this.melFilters[d];for(let f=0;f<c.length;f++)_+=c[f]*u[f];_===0&&(_=1e-30),r[h*this.nfilt+d]=Math.log(_)}}return r}}async function mt(m=Z,o=q){const t=await import(o);return t.env.wasm.wasmPaths=m,t.env.wasm.numThreads=1,t}let dt=null;async function _t(m=st,o=Z,t=q){return dt||(dt=mt(o,t).then(s=>s.InferenceSession.create(m,{executionProviders:["wasm"],graphOptimizationLevel:"all"}))),dt}const It=new pt;class X{constructor(o,t){this._started=!1,this._inferring=!1,this._audioCtx=null,this._stream=null,this._refEmbeddings=new Map,this._lastMatchAt=0,this._lastInferenceAt=0;const{refsStorageKey:s=V,thresholdStorageKey:n=wt,wasmPaths:r=Z,modelPath:a=st,audioProcessorPath:e=gt,ortCdnUrl:h=q}=t||{};this._commands=o,this._refsStorageKey=s,this._thresholdStorageKey=n,this._audioProcessorPath=e,this._wasmPaths=r,this._modelPath=a,this._ortCdnUrl=h;try{const i=localStorage.getItem(this._thresholdStorageKey);this._threshold=i!==null?Math.max(0,Math.min(1,Number(i))):.65}catch{this._threshold=.65}this._initPromise=this._init()}get threshold(){return this._threshold}set threshold(o){this._threshold=Math.max(0,Math.min(1,o));try{localStorage.setItem(this._thresholdStorageKey,String(this._threshold))}catch{}}get listening(){return this._started}async _init(){await _t(this._modelPath,this._wasmPaths,this._ortCdnUrl);const o=new Set;for(const t of this._commands)for(const s of t.triggers)if(!o.has(s.name)&&(o.add(s.name),s.defaultRefPath)){const n=await fetch(s.defaultRefPath);if(n.ok){const r=await n.json();this.addCustomWord(r)}}for(const t of X.loadWords(this._refsStorageKey))this._refEmbeddings.set(t.word_name,t.embeddings);console.info("[Mellon] init complete, loaded refs:",[...this._refEmbeddings.keys()])}async init(){await this._initPromise}addCustomWord(o){if(!(Array.isArray(o.embeddings)&&o.embeddings.length>0))throw new Error("invalid ref file for : "+o.word_name);this._refEmbeddings.set(o.word_name,o.embeddings)}async start(){if(this._started)return;await this._initPromise;let o;try{o=await navigator.mediaDevices.getUserMedia({audio:{noiseSuppression:!1,echoCancellation:!1,autoGainControl:!1,channelCount:1}})}catch{o=await navigator.mediaDevices.getUserMedia({audio:!0})}this._stream=o;const t=new AudioContext({sampleRate:16e3});this._audioCtx=t,await t.audioWorklet.addModule(this._audioProcessorPath);const s=t.createMediaStreamSource(o),n=new AudioWorkletNode(t,"audio-processor");n.port.onmessage=r=>{this._handleBuffer(r.data)},s.connect(n),n.connect(t.destination),this._started=!0}async stop(){if(this._started=!1,this._audioCtx&&(await this._audioCtx.close(),this._audioCtx=null),this._stream){for(const o of this._stream.getTracks())o.stop();this._stream=null}}async _handleBuffer(o){if(this._inferring)return;const t=Date.now();if(!(t-this._lastInferenceAt<300)){this._lastInferenceAt=t,this._inferring=!0;try{const[s,n]=await Promise.all([mt(this._wasmPaths,this._ortCdnUrl),_t(this._modelPath,this._wasmPaths,this._ortCdnUrl)]),r=It.logfbank(o),a=new s.Tensor("float32",r,[1,1,149,64]),e=await n.run({input:a}),h=e[Object.keys(e)[0]].data;let i=!1;for(const l of this._commands){if(i)break;for(const c of l.triggers){const d=this._refEmbeddings.get(c.name);if(!d)continue;const _=this._maxCosineSim(h,d);if(_>=this._threshold&&t-this._lastMatchAt>2e3){this._lastMatchAt=t,console.info(`[Mellon] match: "${c}" sim=${_.toFixed(3)}`),typeof l.onMatch=="function"&&l.onMatch(c.name,_),i=!0;break}}}}catch(s){console.error("[Mellon] inference error:",s)}finally{this._inferring=!1}}}_maxCosineSim(o,t){let s=0;for(const n of t){let r=0;for(let e=0;e<n.length;e++)r+=o[e]*n[e];const a=(r+1)/2;a>s&&(s=a)}return s}static loadWords(o=V){try{const t=localStorage.getItem(o);return t?JSON.parse(t):[]}catch{return[]}}static saveWord(o,t=V){const s=X.loadWords(t).filter(n=>n.word_name!==o.word_name);localStorage.setItem(t,JSON.stringify([...s,o]))}static deleteWord(o,t=V){try{const s=X.loadWords(t).filter(n=>n.word_name!==o);localStorage.setItem(t,JSON.stringify(s))}catch{}}}const Ht=new pt;class Wt{constructor(o,t){this._config={},this._samples=[],this._wordName=o,this._config.modelPath=(t==null?void 0:t.modelPath)||st,this._config.wasmPaths=(t==null?void 0:t.wasmPaths)||Z,this._config.ortCdnUrl=(t==null?void 0:t.ortCdnUrl)||q}async recordSample(){const o=await navigator.mediaDevices.getUserMedia({audio:!0}),t=new AudioContext({sampleRate:16e3}),s=await new Promise((a,e)=>{const h=new MediaRecorder(o),i=[];h.ondataavailable=l=>{l.data.size>0&&i.push(l.data)},h.onstop=async()=>{var l;for(const c of o.getTracks())c.stop();try{const d=await new Blob(i,{type:((l=i[0])==null?void 0:l.type)||"audio/webm"}).arrayBuffer(),_=await t.decodeAudioData(d);await t.close(),a(_.getChannelData(0).slice())}catch(c){e(c)}},h.start(),setTimeout(()=>{try{h.stop()}catch{}},1500)}),n=24e3,r=new Float32Array(n);return r.set(s.slice(0,n)),this._samples.push(r),this._samples.length}deleteSample(o){if(o<0||o>=this._samples.length)throw new RangeError(`index ${o} out of bounds (${this._samples.length} samples)`);return this._samples.splice(o,1),this._samples.length}async generateRef(){const[o,t]=await Promise.all([mt(this._config.wasmPaths,this._config.ortCdnUrl),_t(this._config.modelPath,this._config.wasmPaths,this._config.ortCdnUrl)]),s=[];for(const n of this._samples){const r=Ht.logfbank(n),a=new o.Tensor("float32",r,[1,1,149,64]),e=await t.run({input:a}),h=Array.from(e[Object.keys(e)[0]].data);s.push(h)}return{word_name:this._wordName,model_type:"resnet_50_arc",embeddings:s}}}const Z="https://cdn.jsdelivr.net/npm/onnxruntime-web@1.24.3/dist/",q="https://cdn.jsdelivr.net/npm/onnxruntime-web@1.24.3/dist/ort.wasm.min.mjs",st="https://huggingface.co/ComicScrip/mellon/resolve/main/model.onnx",gt="https://cdn.jsdelivr.net/npm/mellon@0.0.14/dist/assets/audio-processor.js",V="mellon-refs",wt="mellon-threshold";exports.DEFAULT_AUDIO_PROCESSOR_PATH=gt;exports.DEFAULT_MODEL_PATH=st;exports.DEFAULT_ORT_CDN_URL=q;exports.DEFAULT_REFS_STORAGE_KEY=V;exports.DEFAULT_THRESHOLD_STORAGE_KEY=wt;exports.DEFAULT_WASM_PATHS=Z;exports.EnrollmentSession=Wt;exports.Mellon=X;
+"use strict";Object.defineProperty(exports,Symbol.toStringTag,{value:"Module"});function Ut(m){return m&&m.__esModule&&Object.prototype.hasOwnProperty.call(m,"default")?m.default:m}var dt,pt;function Dt(){if(pt)return dt;pt=1;function m(o){if(this.size=o|0,this.size<=1||(this.size&this.size-1)!==0)throw new Error("FFT size must be a power of two and bigger than 1");this._csize=o<<1;for(var t=new Array(this.size*2),s=0;s<t.length;s+=2){const l=Math.PI*s/this.size;t[s]=Math.cos(l),t[s+1]=-Math.sin(l)}this.table=t;for(var n=0,e=1;this.size>e;e<<=1)n++;this._width=n%2===0?n-1:n,this._bitrev=new Array(1<<this._width);for(var c=0;c<this._bitrev.length;c++){this._bitrev[c]=0;for(var r=0;r<this._width;r+=2){var h=this._width-r-2;this._bitrev[c]|=(c>>>r&3)<<h}}this._out=null,this._data=null,this._inv=0}return dt=m,m.prototype.fromComplexArray=function(t,s){for(var n=s||new Array(t.length>>>1),e=0;e<t.length;e+=2)n[e>>>1]=t[e];return n},m.prototype.createComplexArray=function(){const t=new Array(this._csize);for(var s=0;s<t.length;s++)t[s]=0;return t},m.prototype.toComplexArray=function(t,s){for(var n=s||this.createComplexArray(),e=0;e<n.length;e+=2)n[e]=t[e>>>1],n[e+1]=0;return n},m.prototype.completeSpectrum=function(t){for(var s=this._csize,n=s>>>1,e=2;e<n;e+=2)t[s-e]=t[e],t[s-e+1]=-t[e+1]},m.prototype.transform=function(t,s){if(t===s)throw new Error("Input and output buffers must be different");this._out=t,this._data=s,this._inv=0,this._transform4(),this._out=null,this._data=null},m.prototype.realTransform=function(t,s){if(t===s)throw new Error("Input and output buffers must be different");this._out=t,this._data=s,this._inv=0,this._realTransform4(),this._out=null,this._data=null},m.prototype.inverseTransform=function(t,s){if(t===s)throw new Error("Input and output buffers must be different");this._out=t,this._data=s,this._inv=1,this._transform4();for(var n=0;n<t.length;n++)t[n]/=this.size;this._out=null,this._data=null},m.prototype._transform4=function(){var t=this._out,s=this._csize,n=this._width,e=1<<n,c=s/e<<1,r,h,l=this._bitrev;if(c===4)for(r=0,h=0;r<s;r+=c,h++){const u=l[h];this._singleTransform2(r,u,e)}else for(r=0,h=0;r<s;r+=c,h++){const u=l[h];this._singleTransform4(r,u,e)}var i=this._inv?-1:1,a=this.table;for(e>>=2;e>=2;e>>=2){c=s/e<<1;var d=c>>>2;for(r=0;r<s;r+=c)for(var _=r+d,v=r,f=0;v<_;v+=2,f+=e){const u=v,p=u+d,w=p+d,g=w+d,T=t[u],A=t[u+1],b=t[p],y=t[p+1],F=t[w],S=t[w+1],C=t[g],E=t[g+1],M=T,P=A,R=a[f],U=i*a[f+1],D=b*R-y*U,z=b*U+y*R,k=a[2*f],N=i*a[2*f+1],B=F*k-S*N,j=F*N+S*k,K=a[3*f],O=i*a[3*f+1],$=C*K-E*O,G=C*O+E*K,J=M+B,x=P+j,I=M-B,Y=P-j,Q=D+$,L=z+G,W=i*(D-$),V=i*(z-G),tt=J+Q,rt=x+L,ot=J-Q,nt=x-L,at=I+V,it=Y-W,ct=I-V,lt=Y+W;t[u]=tt,t[u+1]=rt,t[p]=at,t[p+1]=it,t[w]=ot,t[w+1]=nt,t[g]=ct,t[g+1]=lt}}},m.prototype._singleTransform2=function(t,s,n){const e=this._out,c=this._data,r=c[s],h=c[s+1],l=c[s+n],i=c[s+n+1],a=r+l,d=h+i,_=r-l,v=h-i;e[t]=a,e[t+1]=d,e[t+2]=_,e[t+3]=v},m.prototype._singleTransform4=function(t,s,n){const e=this._out,c=this._data,r=this._inv?-1:1,h=n*2,l=n*3,i=c[s],a=c[s+1],d=c[s+n],_=c[s+n+1],v=c[s+h],f=c[s+h+1],u=c[s+l],p=c[s+l+1],w=i+v,g=a+f,T=i-v,A=a-f,b=d+u,y=_+p,F=r*(d-u),S=r*(_-p),C=w+b,E=g+y,M=T+S,P=A-F,R=w-b,U=g-y,D=T-S,z=A+F;e[t]=C,e[t+1]=E,e[t+2]=M,e[t+3]=P,e[t+4]=R,e[t+5]=U,e[t+6]=D,e[t+7]=z},m.prototype._realTransform4=function(){var t=this._out,s=this._csize,n=this._width,e=1<<n,c=s/e<<1,r,h,l=this._bitrev;if(c===4)for(r=0,h=0;r<s;r+=c,h++){const ht=l[h];this._singleRealTransform2(r,ht>>>1,e>>>1)}else for(r=0,h=0;r<s;r+=c,h++){const ht=l[h];this._singleRealTransform4(r,ht>>>1,e>>>1)}var i=this._inv?-1:1,a=this.table;for(e>>=2;e>=2;e>>=2){c=s/e<<1;var d=c>>>1,_=d>>>1,v=_>>>1;for(r=0;r<s;r+=c)for(var f=0,u=0;f<=v;f+=2,u+=e){var p=r+f,w=p+_,g=w+_,T=g+_,A=t[p],b=t[p+1],y=t[w],F=t[w+1],S=t[g],C=t[g+1],E=t[T],M=t[T+1],P=A,R=b,U=a[u],D=i*a[u+1],z=y*U-F*D,k=y*D+F*U,N=a[2*u],B=i*a[2*u+1],j=S*N-C*B,K=S*B+C*N,O=a[3*u],$=i*a[3*u+1],G=E*O-M*$,J=E*$+M*O,x=P+j,I=R+K,Y=P-j,Q=R-K,L=z+G,W=k+J,V=i*(z-G),tt=i*(k-J),rt=x+L,ot=I+W,nt=Y+tt,at=Q-V;if(t[p]=rt,t[p+1]=ot,t[w]=nt,t[w+1]=at,f===0){var it=x-L,ct=I-W;t[g]=it,t[g+1]=ct;continue}if(f!==v){var lt=Y,yt=-Q,Tt=x,bt=-I,At=-i*tt,Ft=-i*V,St=-i*W,Ct=-i*L,Et=lt+At,Mt=yt+Ft,Pt=Tt+Ct,Rt=bt-St,ut=r+_-f,vt=r+d-f;t[ut]=Et,t[ut+1]=Mt,t[vt]=Pt,t[vt+1]=Rt}}}},m.prototype._singleRealTransform2=function(t,s,n){const e=this._out,c=this._data,r=c[s],h=c[s+n],l=r+h,i=r-h;e[t]=l,e[t+1]=0,e[t+2]=i,e[t+3]=0},m.prototype._singleRealTransform4=function(t,s,n){const e=this._out,c=this._data,r=this._inv?-1:1,h=n*2,l=n*3,i=c[s],a=c[s+n],d=c[s+h],_=c[s+l],v=i+d,f=i-d,u=a+_,p=r*(a-_),w=v+u,g=f,T=-p,A=v-u,b=f,y=p;e[t]=w,e[t+1]=0,e[t+2]=g,e[t+3]=T,e[t+4]=A,e[t+5]=0,e[t+6]=b,e[t+7]=y},dt}var zt=Dt();const xt=Ut(zt);class ft{constructor(o=16e3,t=512,s=64){this._sampleRate=o,this._nfft=t,this._nfilt=s,this._fft=new xt(t),this._melFilters=this._createMelFilterbank()}_hzToMel(o){return 2595*Math.log10(1+o/700)}_melToHz(o){return 700*(10**(o/2595)-1)}_createMelFilterbank(){const t=this._sampleRate/2,s=this._hzToMel(0),n=this._hzToMel(t),e=new Float32Array(this._nfilt+2);for(let l=0;l<this._nfilt+2;l++)e[l]=s+l*(n-s)/(this._nfilt+1);const r=e.map(l=>this._melToHz(l)).map(l=>Math.floor((this._nfft+1)*l/this._sampleRate)),h=[];for(let l=0;l<this._nfilt;l++){const i=new Float32Array(Math.floor(this._nfft/2)+1);for(let a=r[l];a<r[l+1];a++)i[a]=(a-r[l])/(r[l+1]-r[l]);for(let a=r[l+1];a<r[l+2];a++)i[a]=(r[l+2]-a)/(r[l+2]-r[l+1]);h.push(i)}return h}logfbank(o){const t=Math.floor(.025*this._sampleRate),s=Math.floor(.01*this._sampleRate),n=1+Math.ceil((o.length-t)/s),e=new Float32Array(n*this._nfilt),c=new Float32Array(this._nfft),r=this._fft.createComplexArray();for(let h=0;h<n;h++){const l=h*s;c.fill(0);for(let d=0;d<t&&l+d<o.length;d++)c[d]=o[l+d];const i=this._fft.toComplexArray(c,null);this._fft.transform(r,i);const a=new Float32Array(Math.floor(this._nfft/2)+1);for(let d=0;d<a.length;d++){const _=r[2*d],v=r[2*d+1];a[d]=1/this._nfft*(_*_+v*v),a[d]===0&&(a[d]=1e-30)}for(let d=0;d<this._nfilt;d++){let _=0;const v=this._melFilters[d];for(let f=0;f<a.length;f++)_+=a[f]*v[f];_===0&&(_=1e-30),e[h*this._nfilt+d]=Math.log(_)}}return e}maxCosineSim(o,t){let s=0;for(const n of t){let e=0;for(let r=0;r<n.length;r++)e+=o[r]*n[r];const c=(e+1)/2;c>s&&(s=c)}return s}}async function st(m=Z,o=q){const t=await import(o);return t.env.wasm.wasmPaths=m,t.env.wasm.numThreads=1,t}let _t=null;async function mt(m=et,o=Z,t=q,s){return _t||(_t=st(o,t).then(n=>s?n.InferenceSession.create(new Uint8Array(s),{executionProviders:["wasm"],graphOptimizationLevel:"all"}):n.InferenceSession.create(m,{executionProviders:["wasm"],graphOptimizationLevel:"all"}))),_t}class H{static loadWords(o=X){try{const t=localStorage.getItem(o);return t?JSON.parse(t):[]}catch{return[]}}static saveWord(o,t=X){const s=H.loadWords(t).filter(n=>n.word_name!==o.word_name);localStorage.setItem(t,JSON.stringify([...s,o]))}static deleteWord(o,t=X){try{const s=H.loadWords(t).filter(n=>n.word_name!==o);localStorage.setItem(t,JSON.stringify(s))}catch{}}}class It{constructor(o,t){this._started=!1,this._inferring=!1,this._audioCtx=null,this._stream=null,this._refEmbeddings=new Map,this._lastMatchAt=0,this._lastInferenceAt=0,this._initPromise=null;const{refsStorageKey:s=X,thresholdStorageKey:n=gt,wasmPaths:e=Z,modelPath:c=et,audioProcessorPath:r=wt,ortCdnUrl:h=q,audioUtils:l=new ft}=t||{};this._audioUtils=l,this._commands=o,this._refsStorageKey=s,this._thresholdStorageKey=n,this._audioProcessorPath=r,this._wasmPaths=e,this._modelPath=c,this._ortCdnUrl=h;try{const i=localStorage.getItem(this._thresholdStorageKey);this._threshold=i!==null?Math.max(0,Math.min(1,Number(i))):.65}catch{this._threshold=.65}}get threshold(){return this._threshold}set threshold(o){this._threshold=Math.max(0,Math.min(1,o));try{localStorage.setItem(this._thresholdStorageKey,String(this._threshold))}catch{}}get listening(){return this._started}async _trackFetch(o,t,s){const n=await fetch(o);if(!n.ok)throw new Error(`HTTP ${n.status} fetching ${o}`);const e=Number(n.headers.get("content-length")??"0");if(e>0&&(s.total+=e),!n.body){const a=await n.arrayBuffer();return s.downloaded+=a.byteLength,e||(s.total+=a.byteLength),t==null||t(s.downloaded,s.total),a}const c=n.body.getReader(),r=[];let h=0;for(;;){const{done:a,value:d}=await c.read();if(a)break;r.push(d),h+=d.length,s.downloaded+=d.length,t==null||t(s.downloaded,s.total)}e||(s.total+=h);const l=new Uint8Array(h);let i=0;for(const a of r)l.set(a,i),i+=a.length;return l.buffer}async _init(o){const t={downloaded:0,total:0},s=new Set,n=[];for(const i of this._commands)for(const a of i.triggers)!s.has(a.name)&&a.defaultRefPath&&(s.add(a.name),n.push({name:a.name,path:a.defaultRefPath}));const e=st(this._wasmPaths,this._ortCdnUrl),[c,...r]=await Promise.all([this._trackFetch(this._modelPath,o,t),...n.map(({path:i})=>this._trackFetch(i,o,t))]);await e,await mt(this._modelPath,this._wasmPaths,this._ortCdnUrl,c);const h=H.loadWords(this._refsStorageKey),l=new Set(h.map(i=>i.word_name));for(let i=0;i<n.length;i++)try{const a=JSON.parse(new TextDecoder().decode(r[i]));this.addCustomWord(a),l.has(a.word_name)||H.saveWord(a,this._refsStorageKey)}catch{console.warn(`[Mellon] failed to parse ref file: ${n[i].path}`)}for(const i of h)this._refEmbeddings.set(i.word_name,i.embeddings);console.info("[Mellon] init complete, loaded refs:",[...this._refEmbeddings.keys()])}async init(o){this._initPromise||(this._initPromise=this._init(o)),await this._initPromise}addCustomWord(o){if(!(Array.isArray(o.embeddings)&&o.embeddings.length>0))throw new Error("invalid ref file for : "+o.word_name);this._refEmbeddings.set(o.word_name,o.embeddings)}async start(){if(this._started)return;await this.init();let o;try{o=await navigator.mediaDevices.getUserMedia({audio:{noiseSuppression:!1,echoCancellation:!1,autoGainControl:!1,channelCount:1}})}catch{o=await navigator.mediaDevices.getUserMedia({audio:!0})}this._stream=o;const t=new AudioContext({sampleRate:16e3});this._audioCtx=t,await t.audioWorklet.addModule(this._audioProcessorPath);const s=t.createMediaStreamSource(o),n=new AudioWorkletNode(t,"audio-processor");n.port.onmessage=e=>{this._handleBuffer(e.data)},s.connect(n),n.connect(t.destination),this._started=!0}async stop(){if(this._started=!1,this._audioCtx&&(await this._audioCtx.close(),this._audioCtx=null),this._stream){for(const o of this._stream.getTracks())o.stop();this._stream=null}}async _handleBuffer(o){if(this._inferring)return;const t=Date.now();if(!(t-this._lastInferenceAt<300)){this._lastInferenceAt=t,this._inferring=!0;try{const[s,n]=await Promise.all([st(this._wasmPaths,this._ortCdnUrl),mt(this._modelPath,this._wasmPaths,this._ortCdnUrl)]),e=this._audioUtils.logfbank(o),c=new s.Tensor("float32",e,[1,1,149,64]),r=await n.run({input:c}),h=r[Object.keys(r)[0]].data;let l=!1;for(const i of this._commands){if(l)break;for(const a of i.triggers){const d=this._refEmbeddings.get(a.name);if(!d)continue;const _=this._audioUtils.maxCosineSim(h,d);if(_>=this._threshold&&t-this._lastMatchAt>2e3){this._lastMatchAt=t,console.info(`[Mellon] match: "${a}" sim=${_.toFixed(3)}`),typeof i.onMatch=="function"&&i.onMatch(a.name,_),l=!0;break}}}}catch(s){console.error("[Mellon] inference error:",s)}finally{this._inferring=!1}}}}class Lt{constructor(o,t){this._config={},this._samples=[],this._wordName=o,this._config.modelPath=(t==null?void 0:t.modelPath)||et,this._config.wasmPaths=(t==null?void 0:t.wasmPaths)||Z,this._config.ortCdnUrl=(t==null?void 0:t.ortCdnUrl)||q,this._audioUtils=(t==null?void 0:t.audioUtils)??new ft}async recordSample(){const o=await navigator.mediaDevices.getUserMedia({audio:!0}),t=new AudioContext({sampleRate:16e3}),s=await new Promise((c,r)=>{const h=new MediaRecorder(o),l=[];h.ondataavailable=i=>{i.data.size>0&&l.push(i.data)},h.onstop=async()=>{var i;for(const a of o.getTracks())a.stop();try{const d=await new Blob(l,{type:((i=l[0])==null?void 0:i.type)||"audio/webm"}).arrayBuffer(),_=await t.decodeAudioData(d);await t.close(),c(_.getChannelData(0).slice())}catch(a){r(a)}},h.start(),setTimeout(()=>{try{h.stop()}catch{}},1500)}),n=24e3,e=new Float32Array(n);return e.set(s.slice(0,n)),this._samples.push(e),this._samples.length}deleteSample(o){if(o<0||o>=this._samples.length)throw new RangeError(`index ${o} out of bounds (${this._samples.length} samples)`);return this._samples.splice(o,1),this._samples.length}async generateRef(){const[o,t]=await Promise.all([st(this._config.wasmPaths,this._config.ortCdnUrl),mt(this._config.modelPath,this._config.wasmPaths,this._config.ortCdnUrl)]),s=[];for(const n of this._samples){const e=this._audioUtils.logfbank(n),c=new o.Tensor("float32",e,[1,1,149,64]),r=await t.run({input:c}),h=Array.from(r[Object.keys(r)[0]].data);s.push(h)}return{word_name:this._wordName,model_type:"resnet_50_arc",embeddings:s}}}const Z="https://cdn.jsdelivr.net/npm/onnxruntime-web@1.24.3/dist/",q="https://cdn.jsdelivr.net/npm/onnxruntime-web@1.24.3/dist/ort.wasm.min.mjs",et="https://huggingface.co/ComicScrip/mellon/resolve/main/model.onnx",wt="https://cdn.jsdelivr.net/npm/mellon@0.0.14/dist/assets/audio-processor.js",X="mellon-refs",gt="mellon-threshold";exports.AudioUtils=ft;exports.DEFAULT_AUDIO_PROCESSOR_PATH=wt;exports.DEFAULT_MODEL_PATH=et;exports.DEFAULT_ORT_CDN_URL=q;exports.DEFAULT_REFS_STORAGE_KEY=X;exports.DEFAULT_THRESHOLD_STORAGE_KEY=gt;exports.DEFAULT_WASM_PATHS=Z;exports.Detector=It;exports.EnrollmentSession=Lt;exports.Storage=H;

package/dist/mellon.mjs

@@ -1,217 +1,250 @@
-function Rt(_) {
-  return _ && _.__esModule && Object.prototype.hasOwnProperty.call(_, "default") ? _.default : _;
+function Rt(m) {
+  return m && m.__esModule && Object.prototype.hasOwnProperty.call(m, "default") ? m.default : m;
 }
-var ht, ut;
-function Et() {
-  if (ut) return ht;
-  ut = 1;
-  function _(o) {
+var lt, vt;
+function Ut() {
+  if (vt) return lt;
+  vt = 1;
+  function m(o) {
     if (this.size = o | 0, this.size <= 1 || (this.size & this.size - 1) !== 0)
       throw new Error("FFT size must be a power of two and bigger than 1");
     this._csize = o << 1;
     for (var t = new Array(this.size * 2), s = 0; s < t.length; s += 2) {
-      const i = Math.PI * s / this.size;
-      t[s] = Math.cos(i), t[s + 1] = -Math.sin(i);
+      const h = Math.PI * s / this.size;
+      t[s] = Math.cos(h), t[s + 1] = -Math.sin(h);
     }
     this.table = t;
-    for (var n = 0, r = 1; this.size > r; r <<= 1)
+    for (var n = 0, e = 1; this.size > e; e <<= 1)
       n++;
     this._width = n % 2 === 0 ? n - 1 : n, this._bitrev = new Array(1 << this._width);
-    for (var a = 0; a < this._bitrev.length; a++) {
-      this._bitrev[a] = 0;
-      for (var e = 0; e < this._width; e += 2) {
-        var h = this._width - e - 2;
-        this._bitrev[a] |= (a >>> e & 3) << h;
+    for (var c = 0; c < this._bitrev.length; c++) {
+      this._bitrev[c] = 0;
+      for (var r = 0; r < this._width; r += 2) {
+        var l = this._width - r - 2;
+        this._bitrev[c] |= (c >>> r & 3) << l;
       }
     }
     this._out = null, this._data = null, this._inv = 0;
   }
-  return ht = _, _.prototype.fromComplexArray = function(t, s) {
-    for (var n = s || new Array(t.length >>> 1), r = 0; r < t.length; r += 2)
-      n[r >>> 1] = t[r];
+  return lt = m, m.prototype.fromComplexArray = function(t, s) {
+    for (var n = s || new Array(t.length >>> 1), e = 0; e < t.length; e += 2)
+      n[e >>> 1] = t[e];
     return n;
-  }, _.prototype.createComplexArray = function() {
+  }, m.prototype.createComplexArray = function() {
     const t = new Array(this._csize);
     for (var s = 0; s < t.length; s++)
       t[s] = 0;
     return t;
-  }, _.prototype.toComplexArray = function(t, s) {
-    for (var n = s || this.createComplexArray(), r = 0; r < n.length; r += 2)
-      n[r] = t[r >>> 1], n[r + 1] = 0;
+  }, m.prototype.toComplexArray = function(t, s) {
+    for (var n = s || this.createComplexArray(), e = 0; e < n.length; e += 2)
+      n[e] = t[e >>> 1], n[e + 1] = 0;
     return n;
-  }, _.prototype.completeSpectrum = function(t) {
-    for (var s = this._csize, n = s >>> 1, r = 2; r < n; r += 2)
-      t[s - r] = t[r], t[s - r + 1] = -t[r + 1];
-  }, _.prototype.transform = function(t, s) {
+  }, m.prototype.completeSpectrum = function(t) {
+    for (var s = this._csize, n = s >>> 1, e = 2; e < n; e += 2)
+      t[s - e] = t[e], t[s - e + 1] = -t[e + 1];
+  }, m.prototype.transform = function(t, s) {
     if (t === s)
       throw new Error("Input and output buffers must be different");
     this._out = t, this._data = s, this._inv = 0, this._transform4(), this._out = null, this._data = null;
-  }, _.prototype.realTransform = function(t, s) {
+  }, m.prototype.realTransform = function(t, s) {
     if (t === s)
       throw new Error("Input and output buffers must be different");
     this._out = t, this._data = s, this._inv = 0, this._realTransform4(), this._out = null, this._data = null;
-  }, _.prototype.inverseTransform = function(t, s) {
+  }, m.prototype.inverseTransform = function(t, s) {
     if (t === s)
       throw new Error("Input and output buffers must be different");
     this._out = t, this._data = s, this._inv = 1, this._transform4();
     for (var n = 0; n < t.length; n++)
       t[n] /= this.size;
     this._out = null, this._data = null;
-  }, _.prototype._transform4 = function() {
-    var t = this._out, s = this._csize, n = this._width, r = 1 << n, a = s / r << 1, e, h, i = this._bitrev;
-    if (a === 4)
-      for (e = 0, h = 0; e < s; e += a, h++) {
-        const v = i[h];
-        this._singleTransform2(e, v, r);
+  }, m.prototype._transform4 = function() {
+    var t = this._out, s = this._csize, n = this._width, e = 1 << n, c = s / e << 1, r, l, h = this._bitrev;
+    if (c === 4)
+      for (r = 0, l = 0; r < s; r += c, l++) {
+        const u = h[l];
+        this._singleTransform2(r, u, e);
       }
     else
-      for (e = 0, h = 0; e < s; e += a, h++) {
-        const v = i[h];
-        this._singleTransform4(e, v, r);
+      for (r = 0, l = 0; r < s; r += c, l++) {
+        const u = h[l];
+        this._singleTransform4(r, u, e);
       }
-    var l = this._inv ? -1 : 1, c = this.table;
-    for (r >>= 2; r >= 2; r >>= 2) {
-      a = s / r << 1;
-      var d = a >>> 2;
-      for (e = 0; e < s; e += a)
-        for (var m = e + d, u = e, f = 0; u < m; u += 2, f += r) {
-          const v = u, p = v + d, w = p + d, g = w + d, b = t[v], T = t[v + 1], F = t[p], y = t[p + 1], A = t[w], C = t[w + 1], P = t[g], M = t[g + 1], S = b, R = T, E = c[f], D = l * c[f + 1], z = F * E - y * D, x = F * D + y * E, B = c[2 * f], N = l * c[2 * f + 1], H = A * B - C * N, K = A * N + C * B, k = c[3 * f], $ = l * c[3 * f + 1], G = P * k - M * $, J = P * $ + M * k, L = S + H, U = R + K, I = S - H, Y = R - K, Q = z + G, W = x + J, j = l * (z - G), V = l * (x - J), X = L + Q, st = U + W, rt = L - Q, et = U - W, ot = I + V, nt = Y - j, at = I - V, it = Y + j;
-          t[v] = X, t[v + 1] = st, t[p] = ot, t[p + 1] = nt, t[w] = rt, t[w + 1] = et, t[g] = at, t[g + 1] = it;
+    var i = this._inv ? -1 : 1, a = this.table;
+    for (e >>= 2; e >= 2; e >>= 2) {
+      c = s / e << 1;
+      var d = c >>> 2;
+      for (r = 0; r < s; r += c)
+        for (var _ = r + d, v = r, f = 0; v < _; v += 2, f += e) {
+          const u = v, p = u + d, w = p + d, g = w + d, b = t[u], T = t[u + 1], F = t[p], y = t[p + 1], A = t[w], C = t[w + 1], M = t[g], S = t[g + 1], P = b, R = T, U = a[f], E = i * a[f + 1], D = F * U - y * E, x = F * E + y * U, B = a[2 * f], N = i * a[2 * f + 1], j = A * B - C * N, H = A * N + C * B, K = a[3 * f], $ = i * a[3 * f + 1], L = M * K - S * $, J = M * $ + S * K, G = P + j, z = R + H, I = P - j, Y = R - H, O = D + L, W = x + J, k = i * (D - L), Q = i * (x - J), X = G + O, et = z + W, rt = G - O, ot = z - W, nt = I + Q, at = Y - k, it = I - Q, ct = Y + k;
+          t[u] = X, t[u + 1] = et, t[p] = nt, t[p + 1] = at, t[w] = rt, t[w + 1] = ot, t[g] = it, t[g + 1] = ct;
         }
     }
-  }, _.prototype._singleTransform2 = function(t, s, n) {
-    const r = this._out, a = this._data, e = a[s], h = a[s + 1], i = a[s + n], l = a[s + n + 1], c = e + i, d = h + l, m = e - i, u = h - l;
-    r[t] = c, r[t + 1] = d, r[t + 2] = m, r[t + 3] = u;
-  }, _.prototype._singleTransform4 = function(t, s, n) {
-    const r = this._out, a = this._data, e = this._inv ? -1 : 1, h = n * 2, i = n * 3, l = a[s], c = a[s + 1], d = a[s + n], m = a[s + n + 1], u = a[s + h], f = a[s + h + 1], v = a[s + i], p = a[s + i + 1], w = l + u, g = c + f, b = l - u, T = c - f, F = d + v, y = m + p, A = e * (d - v), C = e * (m - p), P = w + F, M = g + y, S = b + C, R = T - A, E = w - F, D = g - y, z = b - C, x = T + A;
-    r[t] = P, r[t + 1] = M, r[t + 2] = S, r[t + 3] = R, r[t + 4] = E, r[t + 5] = D, r[t + 6] = z, r[t + 7] = x;
-  }, _.prototype._realTransform4 = function() {
-    var t = this._out, s = this._csize, n = this._width, r = 1 << n, a = s / r << 1, e, h, i = this._bitrev;
-    if (a === 4)
-      for (e = 0, h = 0; e < s; e += a, h++) {
-        const ct = i[h];
-        this._singleRealTransform2(e, ct >>> 1, r >>> 1);
+  }, m.prototype._singleTransform2 = function(t, s, n) {
+    const e = this._out, c = this._data, r = c[s], l = c[s + 1], h = c[s + n], i = c[s + n + 1], a = r + h, d = l + i, _ = r - h, v = l - i;
+    e[t] = a, e[t + 1] = d, e[t + 2] = _, e[t + 3] = v;
+  }, m.prototype._singleTransform4 = function(t, s, n) {
+    const e = this._out, c = this._data, r = this._inv ? -1 : 1, l = n * 2, h = n * 3, i = c[s], a = c[s + 1], d = c[s + n], _ = c[s + n + 1], v = c[s + l], f = c[s + l + 1], u = c[s + h], p = c[s + h + 1], w = i + v, g = a + f, b = i - v, T = a - f, F = d + u, y = _ + p, A = r * (d - u), C = r * (_ - p), M = w + F, S = g + y, P = b + C, R = T - A, U = w - F, E = g - y, D = b - C, x = T + A;
+    e[t] = M, e[t + 1] = S, e[t + 2] = P, e[t + 3] = R, e[t + 4] = U, e[t + 5] = E, e[t + 6] = D, e[t + 7] = x;
+  }, m.prototype._realTransform4 = function() {
+    var t = this._out, s = this._csize, n = this._width, e = 1 << n, c = s / e << 1, r, l, h = this._bitrev;
+    if (c === 4)
+      for (r = 0, l = 0; r < s; r += c, l++) {
+        const ht = h[l];
+        this._singleRealTransform2(r, ht >>> 1, e >>> 1);
       }
     else
-      for (e = 0, h = 0; e < s; e += a, h++) {
-        const ct = i[h];
-        this._singleRealTransform4(e, ct >>> 1, r >>> 1);
+      for (r = 0, l = 0; r < s; r += c, l++) {
+        const ht = h[l];
+        this._singleRealTransform4(r, ht >>> 1, e >>> 1);
       }
-    var l = this._inv ? -1 : 1, c = this.table;
-    for (r >>= 2; r >= 2; r >>= 2) {
-      a = s / r << 1;
-      var d = a >>> 1, m = d >>> 1, u = m >>> 1;
-      for (e = 0; e < s; e += a)
-        for (var f = 0, v = 0; f <= u; f += 2, v += r) {
-          var p = e + f, w = p + m, g = w + m, b = g + m, T = t[p], F = t[p + 1], y = t[w], A = t[w + 1], C = t[g], P = t[g + 1], M = t[b], S = t[b + 1], R = T, E = F, D = c[v], z = l * c[v + 1], x = y * D - A * z, B = y * z + A * D, N = c[2 * v], H = l * c[2 * v + 1], K = C * N - P * H, k = C * H + P * N, $ = c[3 * v], G = l * c[3 * v + 1], J = M * $ - S * G, L = M * G + S * $, U = R + K, I = E + k, Y = R - K, Q = E - k, W = x + J, j = B + L, V = l * (x - J), X = l * (B - L), st = U + W, rt = I + j, et = Y + X, ot = Q - V;
-          if (t[p] = st, t[p + 1] = rt, t[w] = et, t[w + 1] = ot, f === 0) {
-            var nt = U - W, at = I - j;
-            t[g] = nt, t[g + 1] = at;
+    var i = this._inv ? -1 : 1, a = this.table;
+    for (e >>= 2; e >= 2; e >>= 2) {
+      c = s / e << 1;
+      var d = c >>> 1, _ = d >>> 1, v = _ >>> 1;
+      for (r = 0; r < s; r += c)
+        for (var f = 0, u = 0; f <= v; f += 2, u += e) {
+          var p = r + f, w = p + _, g = w + _, b = g + _, T = t[p], F = t[p + 1], y = t[w], A = t[w + 1], C = t[g], M = t[g + 1], S = t[b], P = t[b + 1], R = T, U = F, E = a[u], D = i * a[u + 1], x = y * E - A * D, B = y * D + A * E, N = a[2 * u], j = i * a[2 * u + 1], H = C * N - M * j, K = C * j + M * N, $ = a[3 * u], L = i * a[3 * u + 1], J = S * $ - P * L, G = S * L + P * $, z = R + H, I = U + K, Y = R - H, O = U - K, W = x + J, k = B + G, Q = i * (x - J), X = i * (B - G), et = z + W, rt = I + k, ot = Y + X, nt = O - Q;
+          if (t[p] = et, t[p + 1] = rt, t[w] = ot, t[w + 1] = nt, f === 0) {
+            var at = z - W, it = I - k;
+            t[g] = at, t[g + 1] = it;
             continue;
           }
-          if (f !== u) {
-            var it = Y, wt = -Q, gt = U, yt = -I, bt = -l * X, Ft = -l * V, Tt = -l * j, At = -l * W, Ct = it + bt, Pt = wt + Ft, Mt = gt + At, St = yt - Tt, ft = e + m - f, vt = e + d - f;
-            t[ft] = Ct, t[ft + 1] = Pt, t[vt] = Mt, t[vt + 1] = St;
+          if (f !== v) {
+            var ct = Y, wt = -O, gt = z, yt = -I, bt = -i * X, Ft = -i * Q, Tt = -i * k, At = -i * W, Ct = ct + bt, Mt = wt + Ft, St = gt + At, Pt = yt - Tt, ft = r + _ - f, ut = r + d - f;
+            t[ft] = Ct, t[ft + 1] = Mt, t[ut] = St, t[ut + 1] = Pt;
           }
         }
     }
-  }, _.prototype._singleRealTransform2 = function(t, s, n) {
-    const r = this._out, a = this._data, e = a[s], h = a[s + n], i = e + h, l = e - h;
-    r[t] = i, r[t + 1] = 0, r[t + 2] = l, r[t + 3] = 0;
-  }, _.prototype._singleRealTransform4 = function(t, s, n) {
-    const r = this._out, a = this._data, e = this._inv ? -1 : 1, h = n * 2, i = n * 3, l = a[s], c = a[s + n], d = a[s + h], m = a[s + i], u = l + d, f = l - d, v = c + m, p = e * (c - m), w = u + v, g = f, b = -p, T = u - v, F = f, y = p;
-    r[t] = w, r[t + 1] = 0, r[t + 2] = g, r[t + 3] = b, r[t + 4] = T, r[t + 5] = 0, r[t + 6] = F, r[t + 7] = y;
-  }, ht;
+  }, m.prototype._singleRealTransform2 = function(t, s, n) {
+    const e = this._out, c = this._data, r = c[s], l = c[s + n], h = r + l, i = r - l;
+    e[t] = h, e[t + 1] = 0, e[t + 2] = i, e[t + 3] = 0;
+  }, m.prototype._singleRealTransform4 = function(t, s, n) {
+    const e = this._out, c = this._data, r = this._inv ? -1 : 1, l = n * 2, h = n * 3, i = c[s], a = c[s + n], d = c[s + l], _ = c[s + h], v = i + d, f = i - d, u = a + _, p = r * (a - _), w = v + u, g = f, b = -p, T = v - u, F = f, y = p;
+    e[t] = w, e[t + 1] = 0, e[t + 2] = g, e[t + 3] = b, e[t + 4] = T, e[t + 5] = 0, e[t + 6] = F, e[t + 7] = y;
+  }, lt;
 }
-var Dt = Et();
-const zt = /* @__PURE__ */ Rt(Dt);
+var Et = Ut();
+const Dt = /* @__PURE__ */ Rt(Et);
 class pt {
   constructor(o = 16e3, t = 512, s = 64) {
-    this.sampleRate = o, this.nfft = t, this.nfilt = s, this.fft = new zt(t), this.melFilters = this.createMelFilterbank();
+    this._sampleRate = o, this._nfft = t, this._nfilt = s, this._fft = new Dt(t), this._melFilters = this._createMelFilterbank();
   }
-  hzToMel(o) {
+  _hzToMel(o) {
     return 2595 * Math.log10(1 + o / 700);
   }
-  melToHz(o) {
+  _melToHz(o) {
     return 700 * (10 ** (o / 2595) - 1);
   }
-  createMelFilterbank() {
-    const t = this.sampleRate / 2, s = this.hzToMel(0), n = this.hzToMel(t), r = new Float32Array(this.nfilt + 2);
-    for (let i = 0; i < this.nfilt + 2; i++)
-      r[i] = s + i * (n - s) / (this.nfilt + 1);
-    const e = r.map((i) => this.melToHz(i)).map((i) => Math.floor((this.nfft + 1) * i / this.sampleRate)), h = [];
-    for (let i = 0; i < this.nfilt; i++) {
-      const l = new Float32Array(Math.floor(this.nfft / 2) + 1);
-      for (let c = e[i]; c < e[i + 1]; c++)
-        l[c] = (c - e[i]) / (e[i + 1] - e[i]);
-      for (let c = e[i + 1]; c < e[i + 2]; c++)
-        l[c] = (e[i + 2] - c) / (e[i + 2] - e[i + 1]);
-      h.push(l);
+  _createMelFilterbank() {
+    const t = this._sampleRate / 2, s = this._hzToMel(0), n = this._hzToMel(t), e = new Float32Array(this._nfilt + 2);
+    for (let h = 0; h < this._nfilt + 2; h++)
+      e[h] = s + h * (n - s) / (this._nfilt + 1);
+    const r = e.map((h) => this._melToHz(h)).map((h) => Math.floor((this._nfft + 1) * h / this._sampleRate)), l = [];
+    for (let h = 0; h < this._nfilt; h++) {
+      const i = new Float32Array(Math.floor(this._nfft / 2) + 1);
+      for (let a = r[h]; a < r[h + 1]; a++)
+        i[a] = (a - r[h]) / (r[h + 1] - r[h]);
+      for (let a = r[h + 1]; a < r[h + 2]; a++)
+        i[a] = (r[h + 2] - a) / (r[h + 2] - r[h + 1]);
+      l.push(i);
     }
-    return h;
+    return l;
   }
   /** Returns a flat Float32Array of shape [numFrames × nfilt]. */
   logfbank(o) {
-    const t = Math.floor(0.025 * this.sampleRate), s = Math.floor(0.01 * this.sampleRate), n = 1 + Math.ceil((o.length - t) / s), r = new Float32Array(n * this.nfilt), a = new Float32Array(this.nfft), e = this.fft.createComplexArray();
-    for (let h = 0; h < n; h++) {
-      const i = h * s;
-      a.fill(0);
-      for (let d = 0; d < t && i + d < o.length; d++)
-        a[d] = o[i + d];
-      const l = this.fft.toComplexArray(a, null);
-      this.fft.transform(e, l);
-      const c = new Float32Array(Math.floor(this.nfft / 2) + 1);
-      for (let d = 0; d < c.length; d++) {
-        const m = e[2 * d], u = e[2 * d + 1];
-        c[d] = 1 / this.nfft * (m * m + u * u), c[d] === 0 && (c[d] = 1e-30);
+    const t = Math.floor(0.025 * this._sampleRate), s = Math.floor(0.01 * this._sampleRate), n = 1 + Math.ceil((o.length - t) / s), e = new Float32Array(n * this._nfilt), c = new Float32Array(this._nfft), r = this._fft.createComplexArray();
+    for (let l = 0; l < n; l++) {
+      const h = l * s;
+      c.fill(0);
+      for (let d = 0; d < t && h + d < o.length; d++)
+        c[d] = o[h + d];
+      const i = this._fft.toComplexArray(c, null);
+      this._fft.transform(r, i);
+      const a = new Float32Array(Math.floor(this._nfft / 2) + 1);
+      for (let d = 0; d < a.length; d++) {
+        const _ = r[2 * d], v = r[2 * d + 1];
+        a[d] = 1 / this._nfft * (_ * _ + v * v), a[d] === 0 && (a[d] = 1e-30);
       }
-      for (let d = 0; d < this.nfilt; d++) {
-        let m = 0;
-        const u = this.melFilters[d];
-        for (let f = 0; f < c.length; f++)
-          m += c[f] * u[f];
-        m === 0 && (m = 1e-30), r[h * this.nfilt + d] = Math.log(m);
+      for (let d = 0; d < this._nfilt; d++) {
+        let _ = 0;
+        const v = this._melFilters[d];
+        for (let f = 0; f < a.length; f++)
+          _ += a[f] * v[f];
+        _ === 0 && (_ = 1e-30), e[l * this._nfilt + d] = Math.log(_);
       }
     }
-    return r;
+    return e;
+  }
+  maxCosineSim(o, t) {
+    let s = 0;
+    for (const n of t) {
+      let e = 0;
+      for (let r = 0; r < n.length; r++) e += o[r] * n[r];
+      const c = (e + 1) / 2;
+      c > s && (s = c);
+    }
+    return s;
   }
 }
-async function mt(_ = q, o = tt) {
+async function q(m = tt, o = st) {
   const t = await import(
     /* @vite-ignore */
     o
   );
-  return t.env.wasm.wasmPaths = _, t.env.wasm.numThreads = 1, t;
+  return t.env.wasm.wasmPaths = m, t.env.wasm.numThreads = 1, t;
 }
-let lt = null;
-async function dt(_ = _t, o = q, t = tt) {
-  return lt || (lt = mt(o, t).then(
-    (s) => s.InferenceSession.create(_, {
+let dt = null;
+async function _t(m = mt, o = tt, t = st, s) {
+  return dt || (dt = q(o, t).then(
+    (n) => s ? n.InferenceSession.create(new Uint8Array(s), {
+      executionProviders: ["wasm"],
+      graphOptimizationLevel: "all"
+    }) : n.InferenceSession.create(m, {
       executionProviders: ["wasm"],
       graphOptimizationLevel: "all"
     })
-  )), lt;
+  )), dt;
 }
-const xt = new pt();
-class O {
+class V {
+  static loadWords(o = Z) {
+    try {
+      const t = localStorage.getItem(o);
+      return t ? JSON.parse(t) : [];
+    } catch {
+      return [];
+    }
+  }
+  static saveWord(o, t = Z) {
+    const s = V.loadWords(t).filter((n) => n.word_name !== o.word_name);
+    localStorage.setItem(t, JSON.stringify([...s, o]));
+  }
+  static deleteWord(o, t = Z) {
+    try {
+      const s = V.loadWords(t).filter((n) => n.word_name !== o);
+      localStorage.setItem(t, JSON.stringify(s));
+    } catch {
+    }
+  }
+}
+class It {
   constructor(o, t) {
-    this._started = !1, this._inferring = !1, this._audioCtx = null, this._stream = null, this._refEmbeddings = /* @__PURE__ */ new Map(), this._lastMatchAt = 0, this._lastInferenceAt = 0;
+    this._started = !1, this._inferring = !1, this._audioCtx = null, this._stream = null, this._refEmbeddings = /* @__PURE__ */ new Map(), this._lastMatchAt = 0, this._lastInferenceAt = 0, this._initPromise = null;
     const {
       refsStorageKey: s = Z,
-      thresholdStorageKey: n = Wt,
-      wasmPaths: r = q,
-      modelPath: a = _t,
-      audioProcessorPath: e = It,
-      ortCdnUrl: h = tt
+      thresholdStorageKey: n = zt,
+      wasmPaths: e = tt,
+      modelPath: c = mt,
+      audioProcessorPath: r = xt,
+      ortCdnUrl: l = st,
+      audioUtils: h = new pt()
     } = t || {};
-    this._commands = o, this._refsStorageKey = s, this._thresholdStorageKey = n, this._audioProcessorPath = e, this._wasmPaths = r, this._modelPath = a, this._ortCdnUrl = h;
+    this._audioUtils = h, this._commands = o, this._refsStorageKey = s, this._thresholdStorageKey = n, this._audioProcessorPath = r, this._wasmPaths = e, this._modelPath = c, this._ortCdnUrl = l;
     try {
       const i = localStorage.getItem(this._thresholdStorageKey);
       this._threshold = i !== null ? Math.max(0, Math.min(1, Number(i))) : 0.65;
     } catch {
       this._threshold = 0.65;
     }
-    this._initPromise = this._init();
   }
   get threshold() {
     return this._threshold;
@@ -226,25 +259,63 @@ class O {
   get listening() {
     return this._started;
   }
-  async _init() {
-    await dt(this._modelPath, this._wasmPaths, this._ortCdnUrl);
-    const o = /* @__PURE__ */ new Set();
-    for (const t of this._commands)
-      for (const s of t.triggers)
-        if (!o.has(s.name) && (o.add(s.name), s.defaultRefPath)) {
-          const n = await fetch(s.defaultRefPath);
-          if (n.ok) {
-            const r = await n.json();
-            this.addCustomWord(r);
-          }
-        }
-    for (const t of O.loadWords(this._refsStorageKey))
-      this._refEmbeddings.set(t.word_name, t.embeddings);
+  /**
+   * Streams `url`, calling `onProgress(downloaded, total)` after each chunk.
+   * Falls back to a single-shot fetch when the body stream is unavailable.
+   */
+  async _trackFetch(o, t, s) {
+    const n = await fetch(o);
+    if (!n.ok) throw new Error(`HTTP ${n.status} fetching ${o}`);
+    const e = Number(n.headers.get("content-length") ?? "0");
+    if (e > 0 && (s.total += e), !n.body) {
+      const a = await n.arrayBuffer();
+      return s.downloaded += a.byteLength, e || (s.total += a.byteLength), t == null || t(s.downloaded, s.total), a;
+    }
+    const c = n.body.getReader(), r = [];
+    let l = 0;
+    for (; ; ) {
+      const { done: a, value: d } = await c.read();
+      if (a) break;
+      r.push(d), l += d.length, s.downloaded += d.length, t == null || t(s.downloaded, s.total);
+    }
+    e || (s.total += l);
+    const h = new Uint8Array(l);
+    let i = 0;
+    for (const a of r)
+      h.set(a, i), i += a.length;
+    return h.buffer;
+  }
+  async _init(o) {
+    const t = { downloaded: 0, total: 0 }, s = /* @__PURE__ */ new Set(), n = [];
+    for (const i of this._commands)
+      for (const a of i.triggers)
+        !s.has(a.name) && a.defaultRefPath && (s.add(a.name), n.push({ name: a.name, path: a.defaultRefPath }));
+    const e = q(this._wasmPaths, this._ortCdnUrl), [c, ...r] = await Promise.all([
+      this._trackFetch(this._modelPath, o, t),
+      ...n.map(({ path: i }) => this._trackFetch(i, o, t))
+    ]);
+    await e, await _t(this._modelPath, this._wasmPaths, this._ortCdnUrl, c);
+    const l = V.loadWords(this._refsStorageKey), h = new Set(l.map((i) => i.word_name));
+    for (let i = 0; i < n.length; i++)
+      try {
+        const a = JSON.parse(new TextDecoder().decode(r[i]));
+        this.addCustomWord(a), h.has(a.word_name) || V.saveWord(a, this._refsStorageKey);
+      } catch {
+        console.warn(`[Mellon] failed to parse ref file: ${n[i].path}`);
+      }
+    for (const i of l)
+      this._refEmbeddings.set(i.word_name, i.embeddings);
     console.info("[Mellon] init complete, loaded refs:", [...this._refEmbeddings.keys()]);
   }
-  /** Ensures the ONNX model is loaded — call before generateRef() in enrollment. */
-  async init() {
-    await this._initPromise;
+  /**
+   * Loads the ONNX model and all reference embeddings.
+   * Must be called before {@link start}.
+   * Safe to call multiple times — the work is only done once.
+   *
+   * @param onProgress - optional callback invoked as each asset is loaded
+   */
+  async init(o) {
+    this._initPromise || (this._initPromise = this._init(o)), await this._initPromise;
   }
   /** Adds (or replaces) the reference embeddings for a word without restarting. */
   addCustomWord(o) {
@@ -254,7 +325,7 @@ class O {
   }
   async start() {
     if (this._started) return;
-    await this._initPromise;
+    await this.init();
     let o;
     try {
       o = await navigator.mediaDevices.getUserMedia({
@@ -272,8 +343,8 @@ class O {
     const t = new AudioContext({ sampleRate: 16e3 });
     this._audioCtx = t, await t.audioWorklet.addModule(this._audioProcessorPath);
     const s = t.createMediaStreamSource(o), n = new AudioWorkletNode(t, "audio-processor");
-    n.port.onmessage = (r) => {
-      this._handleBuffer(r.data);
+    n.port.onmessage = (e) => {
+      this._handleBuffer(e.data);
     }, s.connect(n), n.connect(t.destination), this._started = !0;
   }
   async stop() {
@@ -288,16 +359,16 @@ class O {
     if (!(t - this._lastInferenceAt < 300)) {
       this._lastInferenceAt = t, this._inferring = !0;
       try {
-        const [s, n] = await Promise.all([mt(this._wasmPaths, this._ortCdnUrl), dt(this._modelPath, this._wasmPaths, this._ortCdnUrl)]), r = xt.logfbank(o), a = new s.Tensor("float32", r, [1, 1, 149, 64]), e = await n.run({ input: a }), h = e[Object.keys(e)[0]].data;
-        let i = !1;
-        for (const l of this._commands) {
-          if (i) break;
-          for (const c of l.triggers) {
-            const d = this._refEmbeddings.get(c.name);
+        const [s, n] = await Promise.all([q(this._wasmPaths, this._ortCdnUrl), _t(this._modelPath, this._wasmPaths, this._ortCdnUrl)]), e = this._audioUtils.logfbank(o), c = new s.Tensor("float32", e, [1, 1, 149, 64]), r = await n.run({ input: c }), l = r[Object.keys(r)[0]].data;
+        let h = !1;
+        for (const i of this._commands) {
+          if (h) break;
+          for (const a of i.triggers) {
+            const d = this._refEmbeddings.get(a.name);
             if (!d) continue;
-            const m = this._maxCosineSim(h, d);
-            if (m >= this._threshold && t - this._lastMatchAt > 2e3) {
-              this._lastMatchAt = t, console.info(`[Mellon] match: "${c}" sim=${m.toFixed(3)}`), typeof l.onMatch == "function" && l.onMatch(c.name, m), i = !0;
+            const _ = this._audioUtils.maxCosineSim(l, d);
+            if (_ >= this._threshold && t - this._lastMatchAt > 2e3) {
+              this._lastMatchAt = t, console.info(`[Mellon] match: "${a}" sim=${_.toFixed(3)}`), typeof i.onMatch == "function" && i.onMatch(a.name, _), h = !0;
               break;
             }
           }
@@ -309,64 +380,34 @@ class O {
       }
     }
   }
-  _maxCosineSim(o, t) {
-    let s = 0;
-    for (const n of t) {
-      let r = 0;
-      for (let e = 0; e < n.length; e++) r += o[e] * n[e];
-      const a = (r + 1) / 2;
-      a > s && (s = a);
-    }
-    return s;
-  }
-  static loadWords(o = Z) {
-    try {
-      const t = localStorage.getItem(o);
-      return t ? JSON.parse(t) : [];
-    } catch {
-      return [];
-    }
-  }
-  static saveWord(o, t = Z) {
-    const s = O.loadWords(t).filter((n) => n.word_name !== o.word_name);
-    localStorage.setItem(t, JSON.stringify([...s, o]));
-  }
-  static deleteWord(o, t = Z) {
-    try {
-      const s = O.loadWords(t).filter((n) => n.word_name !== o);
-      localStorage.setItem(t, JSON.stringify(s));
-    } catch {
-    }
-  }
 }
-const Ut = new pt();
-class jt {
+class Wt {
   constructor(o, t) {
-    this._config = {}, this._samples = [], this._wordName = o, this._config.modelPath = (t == null ? void 0 : t.modelPath) || _t, this._config.wasmPaths = (t == null ? void 0 : t.wasmPaths) || q, this._config.ortCdnUrl = (t == null ? void 0 : t.ortCdnUrl) || tt;
+    this._config = {}, this._samples = [], this._wordName = o, this._config.modelPath = (t == null ? void 0 : t.modelPath) || mt, this._config.wasmPaths = (t == null ? void 0 : t.wasmPaths) || tt, this._config.ortCdnUrl = (t == null ? void 0 : t.ortCdnUrl) || st, this._audioUtils = (t == null ? void 0 : t.audioUtils) ?? new pt();
   }
   /** Records 1.5 s of audio, stores the decoded PCM, returns new sample count. */
   async recordSample() {
-    const o = await navigator.mediaDevices.getUserMedia({ audio: !0 }), t = new AudioContext({ sampleRate: 16e3 }), s = await new Promise((a, e) => {
-      const h = new MediaRecorder(o), i = [];
-      h.ondataavailable = (l) => {
-        l.data.size > 0 && i.push(l.data);
-      }, h.onstop = async () => {
-        var l;
-        for (const c of o.getTracks()) c.stop();
+    const o = await navigator.mediaDevices.getUserMedia({ audio: !0 }), t = new AudioContext({ sampleRate: 16e3 }), s = await new Promise((c, r) => {
+      const l = new MediaRecorder(o), h = [];
+      l.ondataavailable = (i) => {
+        i.data.size > 0 && h.push(i.data);
+      }, l.onstop = async () => {
+        var i;
+        for (const a of o.getTracks()) a.stop();
         try {
-          const d = await new Blob(i, { type: ((l = i[0]) == null ? void 0 : l.type) || "audio/webm" }).arrayBuffer(), m = await t.decodeAudioData(d);
-          await t.close(), a(m.getChannelData(0).slice());
-        } catch (c) {
-          e(c);
+          const d = await new Blob(h, { type: ((i = h[0]) == null ? void 0 : i.type) || "audio/webm" }).arrayBuffer(), _ = await t.decodeAudioData(d);
+          await t.close(), c(_.getChannelData(0).slice());
+        } catch (a) {
+          r(a);
         }
-      }, h.start(), setTimeout(() => {
+      }, l.start(), setTimeout(() => {
         try {
-          h.stop();
+          l.stop();
         } catch {
         }
       }, 1500);
-    }), n = 24e3, r = new Float32Array(n);
-    return r.set(s.slice(0, n)), this._samples.push(r), this._samples.length;
+    }), n = 24e3, e = new Float32Array(n);
+    return e.set(s.slice(0, n)), this._samples.push(e), this._samples.length;
   }
   /** Removes the sample at the given index. Returns the new sample count. */
   deleteSample(o) {
@@ -376,22 +417,24 @@ class jt {
   }
   /** Runs ONNX inference on every recorded sample to produce reference embeddings. */
   async generateRef() {
-    const [o, t] = await Promise.all([mt(this._config.wasmPaths, this._config.ortCdnUrl), dt(this._config.modelPath, this._config.wasmPaths, this._config.ortCdnUrl)]), s = [];
+    const [o, t] = await Promise.all([q(this._config.wasmPaths, this._config.ortCdnUrl), _t(this._config.modelPath, this._config.wasmPaths, this._config.ortCdnUrl)]), s = [];
     for (const n of this._samples) {
-      const r = Ut.logfbank(n), a = new o.Tensor("float32", r, [1, 1, 149, 64]), e = await t.run({ input: a }), h = Array.from(e[Object.keys(e)[0]].data);
-      s.push(h);
+      const e = this._audioUtils.logfbank(n), c = new o.Tensor("float32", e, [1, 1, 149, 64]), r = await t.run({ input: c }), l = Array.from(r[Object.keys(r)[0]].data);
+      s.push(l);
     }
     return { word_name: this._wordName, model_type: "resnet_50_arc", embeddings: s };
   }
 }
-const q = "https://cdn.jsdelivr.net/npm/onnxruntime-web@1.24.3/dist/", tt = "https://cdn.jsdelivr.net/npm/onnxruntime-web@1.24.3/dist/ort.wasm.min.mjs", _t = "https://huggingface.co/ComicScrip/mellon/resolve/main/model.onnx", It = "https://cdn.jsdelivr.net/npm/mellon@0.0.14/dist/assets/audio-processor.js", Z = "mellon-refs", Wt = "mellon-threshold";
+const tt = "https://cdn.jsdelivr.net/npm/onnxruntime-web@1.24.3/dist/", st = "https://cdn.jsdelivr.net/npm/onnxruntime-web@1.24.3/dist/ort.wasm.min.mjs", mt = "https://huggingface.co/ComicScrip/mellon/resolve/main/model.onnx", xt = "https://cdn.jsdelivr.net/npm/mellon@0.0.14/dist/assets/audio-processor.js", Z = "mellon-refs", zt = "mellon-threshold";
 export {
-  It as DEFAULT_AUDIO_PROCESSOR_PATH,
-  _t as DEFAULT_MODEL_PATH,
-  tt as DEFAULT_ORT_CDN_URL,
+  pt as AudioUtils,
+  xt as DEFAULT_AUDIO_PROCESSOR_PATH,
+  mt as DEFAULT_MODEL_PATH,
+  st as DEFAULT_ORT_CDN_URL,
   Z as DEFAULT_REFS_STORAGE_KEY,
-  Wt as DEFAULT_THRESHOLD_STORAGE_KEY,
-  q as DEFAULT_WASM_PATHS,
-  jt as EnrollmentSession,
-  O as Mellon
+  zt as DEFAULT_THRESHOLD_STORAGE_KEY,
+  tt as DEFAULT_WASM_PATHS,
+  It as Detector,
+  Wt as EnrollmentSession,
+  V as Storage
 };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mellon",
-  "version": "0.0.18",
+  "version": "0.0.20",
   "description": "Offline, in-browser voice commands powered by EfficientWord-Net (ResNet-50 ArcFace).",
   "type": "module",
   "main": "./dist/mellon.cjs",