mellon 0.0.3 → 0.0.5

package/README.md

@@ -1,12 +1,11 @@
 # mellon
 
-Offline, fully in-browser **hotword / wake-word detection** powered by [EfficientWord-Net](https://github.com/Ant-Brain/EfficientWord-Net) (ResNet-50 ArcFace). Works as a zero-dependency npm library or as a standalone PWA.
+Offline, fully in-browser **hotword / wake-word detection** powered by [EfficientWord-Net](https://github.com/Ant-Brain/EfficientWord-Net) (ResNet-50 ArcFace).
 
 - **100% offline** — ONNX inference runs in the browser via WebAssembly; no server, no cloud.
 - **Speaker-independent** — the model generalises across voices out of the box.
-- **Custom words** — enroll any phrase with ≥ 3 audio samples; no retraining.
+- **Custom words** — enroll any phrase with ≥ 3 audio samples.
 - **TypeScript-ready** — ships with full `.d.ts` declarations.
-- **Tiny API surface** — one class, zero config.
 
 ---
 
@@ -81,16 +80,14 @@ Refs are fetched automatically during `start()`. You can enroll your own words
 By default, the WASM runtime and model load from the jsDelivr CDN — no setup needed. For air-gapped or private-network deployments, copy the assets locally and tell the library where to find them:
 
 ```bash
-cp -r node_modules/mellon/dist/wasm    public/mellon-assets/wasm
-cp    node_modules/mellon/dist/models/model.onnx  public/mellon-assets/model.onnx
+cp -r node_modules/mellon/dist/assets    public/mellon-assets/
 ```
 
 Then pass the paths to the constructor:
 
 ```js
 new Mellon({
-  wasmBasePath: '/mellon-assets/wasm/',   // trailing slash required
-  modelUrl:     '/mellon-assets/model.onnx',
+  assetsPath: '/mellon-assets',   // trailing slash required
 })
 ```
 
@@ -104,8 +101,7 @@ export default {
   plugins: [
     viteStaticCopy({
       targets: [
-        { src: 'node_modules/mellon/dist/wasm/*',           dest: 'mellon-assets/wasm' },
-        { src: 'node_modules/mellon/dist/models/model.onnx', dest: 'mellon-assets' },
+        { src: 'node_modules/mellon/dist/assets/*', dest: 'mellon-assets' }
       ],
     }),
   ],

package/dist/assets/audio-processor.js

@@ -0,0 +1,37 @@
+/**
+ * public/audio-processor.js
+ * AudioWorklet that runs at 16 kHz and continuously emits the last
+ * 1.5-second window (24 000 samples) via a circular buffer.
+ *
+ * The main thread receives a fresh Float32Array on every
+ * AudioWorklet quantum (128 samples ≈ every 8 ms at 16 kHz).
+ * The inference loop in engine.js rate-limits to avoid excessive work.
+ */
+class AudioProcessor extends AudioWorkletProcessor {
+    constructor() {
+        super()
+        this._size = 24000  // 1.5 s × 16 000 Hz
+        this._buf = new Float32Array(this._size)
+        this._ptr = 0
+    }
+
+    process(inputs) {
+        const ch = inputs[0]?.[0]
+        if (!ch) return true
+
+        for (let i = 0; i < ch.length; i++) {
+            this._buf[this._ptr] = ch[i]
+            this._ptr = (this._ptr + 1) % this._size
+        }
+
+        // Send an ordered copy of the ring buffer
+        const out = new Float32Array(this._size)
+        for (let i = 0; i < this._size; i++) {
+            out[i] = this._buf[(this._ptr + i) % this._size]
+        }
+        this.port.postMessage(out)
+        return true
+    }
+}
+
+registerProcessor('audio-processor', AudioProcessor)

package/dist/assets/manifest.json

@@ -0,0 +1,16 @@
+{
+  "name": "Mellon STT",
+  "short_name": "Mellon",
+  "description": "Offline hotword detection demo — EfficientWord-Net in the browser",
+  "start_url": "./",
+  "display": "standalone",
+  "background_color": "#080810",
+  "theme_color": "#080810",
+  "icons": [
+    {
+      "src": "data:image/svg+xml,<svg xmlns='http://www.w3.org/2000/svg' viewBox='0 0 100 100'><text y='.9em' font-size='90'>🎙</text></svg>",
+      "sizes": "any",
+      "type": "image/svg+xml"
+    }
+  ]
+}

package/dist/assets/sw.js

@@ -0,0 +1,76 @@
+/**
+ * public/sw.js — Service Worker for offline-first caching.
+ *
+ * Strategy:
+ *   • model.onnx, *.wasm, *_ref.json  → Cache-first (immutable large assets)
+ *   • JS / CSS / HTML                 → Stale-while-revalidate
+ *   • Everything else                 → Network-first with cache fallback
+ */
+
+const STATIC_CACHE = 'mellon-static-v1'
+const MODEL_CACHE = 'mellon-model-v1'
+
+// ─── Lifecycle ───────────────────────────────────────────────────────────────
+
+self.addEventListener('install', () => self.skipWaiting())
+
+self.addEventListener('activate', e => {
+    e.waitUntil(
+        caches.keys().then(keys =>
+            Promise.all(
+                keys
+                    .filter(k => k !== STATIC_CACHE && k !== MODEL_CACHE)
+                    .map(k => caches.delete(k))
+            )
+        ).then(() => self.clients.claim())
+    )
+})
+
+// ─── Fetch ───────────────────────────────────────────────────────────────────
+
+self.addEventListener('fetch', e => {
+    const { request } = e
+    const url = new URL(request.url)
+    const p = url.pathname
+
+    // Large immutable assets: model, wasm files, reference JSONs
+    if (p.endsWith('.onnx') || p.endsWith('.wasm') || p.includes('_ref.json') || p.endsWith('.mjs')) {
+        e.respondWith(cacheFirst(request, MODEL_CACHE))
+        return
+    }
+
+    // App shell: stale-while-revalidate
+    if (request.mode === 'navigate' || p.endsWith('.js') || p.endsWith('.css')) {
+        e.respondWith(staleWhileRevalidate(request, STATIC_CACHE))
+        return
+    }
+
+    // Default: network first
+    e.respondWith(
+        fetch(request).catch(() => caches.match(request))
+    )
+})
+
+// ─── Helpers ─────────────────────────────────────────────────────────────────
+
+async function cacheFirst(request, cacheName) {
+    const cache = await caches.open(cacheName)
+    const cached = await cache.match(request)
+    if (cached) return cached
+
+    const response = await fetch(request)
+    if (response.ok) cache.put(request, response.clone())
+    return response
+}
+
+async function staleWhileRevalidate(request, cacheName) {
+    const cache = await caches.open(cacheName)
+    const cached = await cache.match(request)
+
+    const fetchPromise = fetch(request).then(r => {
+        if (r.ok) cache.put(request, r.clone())
+        return r
+    }).catch(() => null)
+
+    return cached || fetchPromise
+}

package/dist/index.d.ts

@@ -0,0 +1,295 @@
+// mellon type declarations
+
+// ─── Shared data types ───────────────────────────────────────────────────────
+
+export interface RefData {
+    word_name: string
+    model_type: 'resnet_50_arc'
+    embeddings: number[][]
+}
+
+export interface MatchEventDetail {
+    /** The detected word name. */
+    name: string
+    /** Similarity score that triggered detection (0–1). */
+    confidence: number
+    /** Unix timestamp (ms) of the detection. */
+    timestamp: number
+}
+
+export interface SampleInfo {
+    audioBuffer: Float32Array
+    name: string
+}
+
+// ─── Engine ──────────────────────────────────────────────────────────────────
+
+export interface EngineConfig {
+    /**
+     * Base URL where ORT WASM files are served (trailing slash required).
+     * Defaults to the jsDelivr CDN. Override for offline / intranet use.
+     * @example '/mellon-assets/wasm/'
+     */
+    wasmBasePath?: string
+    /**
+     * Full URL to model.onnx.
+     * Defaults to the jsDelivr CDN. Override for offline / intranet use.
+     * @example '/mellon-assets/model.onnx'
+     */
+    modelUrl?: string
+}
+
+/**
+ * Override asset paths. Optional — by default assets load from the jsDelivr CDN.
+ * Call this before loadModel() when deploying offline or on a private network.
+ */
+export function configure(config: EngineConfig): void
+
+/**
+ * Load (or return the already-loaded) ONNX inference session.
+ * Idempotent — safe to call multiple times.
+ *
+ * @param onProgress  Called with values 0.0 → 1.0 as the model downloads.
+ */
+export function loadModel(onProgress?: (progress: number) => void): Promise<void>
+
+/**
+ * Compute a 256-dim L2-normalised embedding from a log-mel spectrogram.
+ * Requires loadModel() to have completed first.
+ *
+ * @param spectrogram  Flat Float32Array of shape [149 × 64] from logfbank().
+ */
+export function embed(spectrogram: Float32Array): Promise<Float32Array>
+
+// ─── Mel feature extraction ───────────────────────────────────────────────────
+
+/**
+ * Compute a log-mel spectrogram from a 1.5-second 16 kHz audio buffer.
+ *
+ * @param signal  24 000 samples at 16 kHz (1.5 seconds).
+ * @returns       Float32Array of shape [149 × 64] (frames × mel-bins).
+ */
+export function logfbank(signal: Float32Array): Float32Array
+
+// ─── Similarity helpers ───────────────────────────────────────────────────────
+
+/**
+ * Cosine similarity normalised to [0, 1].
+ * Assumes both vectors are L2-normalised (as the ArcFace model guarantees).
+ */
+export function cosineSim(a: Float32Array | number[], b: Float32Array | number[]): number
+
+/**
+ * Maximum cosine similarity between `embedding` and any of `refs`.
+ */
+export function maxSimilarity(
+    embedding: Float32Array,
+    refs: number[][] | Float32Array[],
+): number
+
+// ─── HotwordDetector ─────────────────────────────────────────────────────────
+
+export interface DetectorOptions {
+    /** Human-readable label for this word. */
+    name: string
+    /** Reference embeddings (N × 256), e.g. from a RefData.embeddings array. */
+    refEmbeddings: number[][] | Float32Array[]
+    /** Detection threshold in [0, 1]. Default: 0.65 */
+    threshold?: number
+    /** Minimum milliseconds between successive 'match' events. Default: 2000 */
+    relaxationMs?: number
+    /** Minimum milliseconds between consecutive inference runs. Default: 300 */
+    inferenceGapMs?: number
+}
+
+/**
+ * Stateful detector for a single hotword.
+ *
+ * @example
+ *   const myRef = await importRefFile(file)  // or load from your source
+ *   const d = new HotwordDetector({ name: myRef.word_name, refEmbeddings: myRef.embeddings })
+ *   d.addEventListener('match', e => console.log(e.detail.name, e.detail.confidence))
+ *   // In AudioWorklet onmessage handler:
+ *   const score = await d.scoreFrame(audioBuffer)
+ */
+export class HotwordDetector extends EventTarget {
+    constructor(opts: DetectorOptions)
+
+    readonly name: string
+    /** Most recent similarity score (0–1). */
+    readonly lastScore: number
+    threshold: number
+    relaxationMs: number
+    inferenceGapMs: number
+    refEmbeddings: number[][] | Float32Array[]
+
+    /**
+     * Score a 1.5-second audio frame. Rate-limited to inferenceGapMs.
+     *
+     * @param audioBuffer  24 000 samples at 16 kHz.
+     * @returns            Similarity score, or null when rate-limited.
+     */
+    scoreFrame(audioBuffer: Float32Array): Promise<number | null>
+
+    addEventListener(
+        type: 'match',
+        listener: (event: CustomEvent<MatchEventDetail>) => void,
+        options?: boolean | AddEventListenerOptions,
+    ): void
+    addEventListener(
+        type: string,
+        listener: EventListenerOrEventListenerObject,
+        options?: boolean | AddEventListenerOptions,
+    ): void
+}
+
+// ─── EnrollmentSession ────────────────────────────────────────────────────────
+
+/**
+ * Manages recording and embedding generation for a custom wake word.
+ *
+ * @example
+ *   const session = new EnrollmentSession('hello')
+ *   await session.recordSample()   // record 1.5 s from mic (repeat 3+ times)
+ *   const ref = await session.generateRef()
+ *   saveCustomRef(ref)             // persist to localStorage
+ */
+export class EnrollmentSession extends EventTarget {
+    constructor(wordName: string)
+
+    readonly wordName: string
+    readonly sampleCount: number
+    readonly samples: SampleInfo[]
+
+    /** Record exactly 1.5 seconds from the microphone. Returns 1-based sample index. */
+    recordSample(): Promise<number>
+
+    /** Decode an uploaded audio File and add it as a sample. Returns 1-based index. */
+    addAudioFile(file: File): Promise<number>
+
+    /** Remove a sample by 0-based index. */
+    removeSample(idx: number): void
+
+    /** Remove all samples. */
+    clearSamples(): void
+
+    /**
+     * Generate reference embeddings from the accumulated samples (minimum 3 required).
+     * Returns a RefData object ready to pass to saveCustomRef() or Mellon.addCustomWord().
+     */
+    generateRef(): Promise<RefData>
+
+    addEventListener(type: 'recording-start', listener: (event: CustomEvent) => void, options?: boolean | AddEventListenerOptions): void
+    addEventListener(type: 'sample-added', listener: (event: CustomEvent<{ count: number; name: string }>) => void, options?: boolean | AddEventListenerOptions): void
+    addEventListener(type: 'samples-changed', listener: (event: CustomEvent<{ count: number }>) => void, options?: boolean | AddEventListenerOptions): void
+    addEventListener(type: 'generating', listener: (event: CustomEvent<{ total: number }>) => void, options?: boolean | AddEventListenerOptions): void
+    addEventListener(type: 'progress', listener: (event: CustomEvent<{ done: number; total: number }>) => void, options?: boolean | AddEventListenerOptions): void
+    addEventListener(type: string, listener: EventListenerOrEventListenerObject, options?: boolean | AddEventListenerOptions): void
+}
+
+// ─── Mellon (high-level API) ─────────────────────────────────────────────────
+
+export interface MellonOptions {
+    /** Words to detect. Refs must be registered via `addCustomWord()` or `refs` before `start()`. */
+    words?: string[]
+    /**
+     * Reference data to preload during `init()`. Each entry is either:
+     * - a URL string pointing to a hosted `_ref.json` file, or
+     * - an inline `RefData` object.
+     *
+     * @example
+     *   refs: [
+     *     'https://example.com/hello_ref.json',
+     *     'https://example.com/stop_ref.json',
+     *   ]
+     */
+    refs?: (string | RefData)[]
+    /** Detection threshold [0, 1]. Default: 0.65 */
+    threshold?: number
+    /** Minimum milliseconds between successive match events per word. Default: 2000 */
+    relaxationMs?: number
+    /** Minimum milliseconds between consecutive inference runs. Default: 300 */
+    inferenceGapMs?: number
+    /**
+     * Override the ORT WASM base URL. Defaults to the jsDelivr CDN.
+     * Only needed for offline / intranet deployments (trailing slash required).
+     * @example '/mellon-assets/wasm/'
+     */
+    wasmBasePath?: string
+    /**
+     * Override the model.onnx URL. Defaults to the jsDelivr CDN.
+     * Only needed for offline / intranet deployments.
+     * @example '/mellon-assets/model.onnx'
+     */
+    modelUrl?: string
+}
+
+/**
+ * High-level, all-in-one hotword detector.
+ *
+ * @example
+ *   const stt = new Mellon({
+ *     wasmBasePath: '/assets/wasm/',
+ *     modelUrl: '/assets/model.onnx',
+ *   })
+ *   await stt.init(pct => progressBar.style.width = pct * 100 + '%')
+ *   await stt.start()
+ *   stt.addEventListener('match', e => console.log(e.detail.name, e.detail.confidence))
+ */
+export class Mellon extends EventTarget {
+    constructor(opts?: MellonOptions)
+
+    /** True after init() has completed successfully. */
+    readonly isInitialized: boolean
+
+    /** True while start() is active (microphone is open). */
+    readonly isRunning: boolean
+
+    /**
+     * Load the ONNX model and cache built-in reference embeddings.
+     * Optional — start() auto-calls init() when needed.
+     *
+     * @param onProgress  Progress callback, 0.0 → 1.0.
+     */
+    init(onProgress?: (progress: number) => void): Promise<void>
+
+    /**
+     * Request microphone access and start hotword detection.
+     * Resolves once audio pipeline is running.
+     *
+     * @param words  Optional subset of words to activate (must have refs loaded).
+     */
+    start(words?: string[]): Promise<void>
+
+    /** Stop detection and release the microphone + AudioContext. */
+    stop(): void
+
+    /**
+     * Register reference embeddings for a word.
+     * Can be called before or after start().
+     */
+    addCustomWord(refData: RefData): void
+
+    /**
+     * Create an EnrollmentSession for recording a new custom word.
+     * Call addCustomWord() with the result of session.generateRef().
+     */
+    enrollWord(wordName: string): EnrollmentSession
+
+    /** Return all custom word refs stored in localStorage. */
+    static loadWords(): RefData[]
+    /** Persist a word ref to localStorage (replaces any existing entry with the same name). */
+    static saveWord(refData: RefData): void
+    /** Delete a word ref from localStorage by name. */
+    static deleteWord(wordName: string): void
+    /** Parse an uploaded ref JSON file. */
+    static importWordFile(file: File): Promise<RefData>
+    /** Trigger a browser download of a ref as a JSON file. */
+    static exportWord(refData: RefData): void
+
+    addEventListener(type: 'match', listener: (event: CustomEvent<MatchEventDetail>) => void, options?: boolean | AddEventListenerOptions): void
+    addEventListener(type: 'ready', listener: (event: CustomEvent) => void, options?: boolean | AddEventListenerOptions): void
+    addEventListener(type: 'error', listener: (event: CustomEvent<{ error: Error }>) => void, options?: boolean | AddEventListenerOptions): void
+    addEventListener(type: string, listener: EventListenerOrEventListenerObject, options?: boolean | AddEventListenerOptions): void
+}
+

package/dist/mellon.cjs

@@ -0,0 +1,38 @@
+"use strict";Object.defineProperty(exports,Symbol.toStringTag,{value:"Module"});const Jt="0.0.5",Kt=[1,1,149,64],Yt=`https://cdn.jsdelivr.net/npm/mellon@${Jt}/dist/assets`,st={assetsPath:`${Yt}`};let I=null,q=null,tt=null;function Qt({assetsPath:i}={}){i!==void 0&&(st.assetsPath=i),I=null,q=null,tt=null}async function Vt(i){return I?(i==null||i(1),I):q||(q=(async()=>{const n=st.assetsPath.endsWith("/")?st.assetsPath:st.assetsPath+"/",t=n+"ort.all.min.mjs",e=n+"model.onnx";tt=await new Function("url","return import(url)")(t),tt.env.wasm.wasmPaths=n;const s=await fetch(e);if(!s.ok)throw new Error(`Failed to fetch model: ${s.status}`);const a=parseInt(s.headers.get("content-length")||"0",10),r=s.body.getReader(),c=[];let l=0;for(;;){const{done:m,value:_}=await r.read();if(m)break;c.push(_),l+=_.byteLength,a>0&&(i==null||i(l/a))}const h=new Uint8Array(l);let d=0;for(const m of c)h.set(m,d),d+=m.byteLength;return I=await tt.InferenceSession.create(h.buffer,{executionProviders:["wasm"],graphOptimizationLevel:"all"}),i==null||i(1),I})(),q)}async function St(i){if(!I)throw new Error("Model not loaded — call loadModel() first");const n=new tt.Tensor("float32",i,Kt),t=await I.run({input:n}),e=Object.keys(t)[0];return t[e].data}function Xt(i){return i&&i.__esModule&&Object.prototype.hasOwnProperty.call(i,"default")?i.default:i}var mt,Et;function Zt(){if(Et)return mt;Et=1;function i(n){if(this.size=n|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=n<<1;for(var t=new Array(this.size*2),e=0;e<t.length;e+=2){const l=Math.PI*e/this.size;t[e]=Math.cos(l),t[e+1]=-Math.sin(l)}this.table=t;for(var o=0,s=1;this.size>s;s<<=1)o++;this._width=o%2===0?o-1:o,this._bitrev=new Array(1<<this._width);for(var a=0;a<this._bitrev.length;a++){this._bitrev[a]=0;for(var r=0;r<this._width;r+=2){var c=this._width-r-2;this._bitrev[a]|=(a>>>r&3)<<c}}this._out=null,this._data=null,this._inv=0}return mt=i,i.prototype.fromComplexArray=function(t,e){for(var o=e||new Array(t.length>>>1),s=0;s<t.length;s+=2)o[s>>>1]=t[s];return o},i.prototype.createComplexArray=function(){const t=new Array(this._csize);for(var e=0;e<t.length;e++)t[e]=0;return t},i.prototype.toComplexArray=function(t,e){for(var o=e||this.createComplexArray(),s=0;s<o.length;s+=2)o[s]=t[s>>>1],o[s+1]=0;return o},i.prototype.completeSpectrum=function(t){for(var e=this._csize,o=e>>>1,s=2;s<o;s+=2)t[e-s]=t[s],t[e-s+1]=-t[s+1]},i.prototype.transform=function(t,e){if(t===e)throw new Error("Input and output buffers must be different");this._out=t,this._data=e,this._inv=0,this._transform4(),this._out=null,this._data=null},i.prototype.realTransform=function(t,e){if(t===e)throw new Error("Input and output buffers must be different");this._out=t,this._data=e,this._inv=0,this._realTransform4(),this._out=null,this._data=null},i.prototype.inverseTransform=function(t,e){if(t===e)throw new Error("Input and output buffers must be different");this._out=t,this._data=e,this._inv=1,this._transform4();for(var o=0;o<t.length;o++)t[o]/=this.size;this._out=null,this._data=null},i.prototype._transform4=function(){var t=this._out,e=this._csize,o=this._width,s=1<<o,a=e/s<<1,r,c,l=this._bitrev;if(a===4)for(r=0,c=0;r<e;r+=a,c++){const u=l[c];this._singleTransform2(r,u,s)}else for(r=0,c=0;r<e;r+=a,c++){const u=l[c];this._singleTransform4(r,u,s)}var h=this._inv?-1:1,d=this.table;for(s>>=2;s>=2;s>>=2){a=e/s<<1;var m=a>>>2;for(r=0;r<e;r+=a)for(var _=r+m,g=r,f=0;g<_;g+=2,f+=s){const u=g,p=u+m,v=p+m,w=v+m,b=t[u],A=t[u+1],E=t[p],y=t[p+1],F=t[v],M=t[v+1],C=t[w],T=t[w+1],x=b,R=A,z=d[f],S=h*d[f+1],N=E*z-y*S,k=E*S+y*z,P=d[2*f],L=h*d[2*f+1],G=F*P-M*L,H=F*L+M*P,J=d[3*f],K=h*d[3*f+1],Y=C*J-T*K,Q=C*K+T*J,V=x+G,W=R+H,B=x-G,X=R-H,Z=N+Y,U=k+Q,$=h*(N-Y),O=h*(k-Q),et=V+Z,ot=W+U,at=V-Z,it=W-U,ct=B+O,lt=X-$,ht=B-O,dt=X+$;t[u]=et,t[u+1]=ot,t[p]=ct,t[p+1]=lt,t[v]=at,t[v+1]=it,t[w]=ht,t[w+1]=dt}}},i.prototype._singleTransform2=function(t,e,o){const s=this._out,a=this._data,r=a[e],c=a[e+1],l=a[e+o],h=a[e+o+1],d=r+l,m=c+h,_=r-l,g=c-h;s[t]=d,s[t+1]=m,s[t+2]=_,s[t+3]=g},i.prototype._singleTransform4=function(t,e,o){const s=this._out,a=this._data,r=this._inv?-1:1,c=o*2,l=o*3,h=a[e],d=a[e+1],m=a[e+o],_=a[e+o+1],g=a[e+c],f=a[e+c+1],u=a[e+l],p=a[e+l+1],v=h+g,w=d+f,b=h-g,A=d-f,E=m+u,y=_+p,F=r*(m-u),M=r*(_-p),C=v+E,T=w+y,x=b+M,R=A-F,z=v-E,S=w-y,N=b-M,k=A+F;s[t]=C,s[t+1]=T,s[t+2]=x,s[t+3]=R,s[t+4]=z,s[t+5]=S,s[t+6]=N,s[t+7]=k},i.prototype._realTransform4=function(){var t=this._out,e=this._csize,o=this._width,s=1<<o,a=e/s<<1,r,c,l=this._bitrev;if(a===4)for(r=0,c=0;r<e;r+=a,c++){const ut=l[c];this._singleRealTransform2(r,ut>>>1,s>>>1)}else for(r=0,c=0;r<e;r+=a,c++){const ut=l[c];this._singleRealTransform4(r,ut>>>1,s>>>1)}var h=this._inv?-1:1,d=this.table;for(s>>=2;s>=2;s>>=2){a=e/s<<1;var m=a>>>1,_=m>>>1,g=_>>>1;for(r=0;r<e;r+=a)for(var f=0,u=0;f<=g;f+=2,u+=s){var p=r+f,v=p+_,w=v+_,b=w+_,A=t[p],E=t[p+1],y=t[v],F=t[v+1],M=t[w],C=t[w+1],T=t[b],x=t[b+1],R=A,z=E,S=d[u],N=h*d[u+1],k=y*S-F*N,P=y*N+F*S,L=d[2*u],G=h*d[2*u+1],H=M*L-C*G,J=M*G+C*L,K=d[3*u],Y=h*d[3*u+1],Q=T*K-x*Y,V=T*Y+x*K,W=R+H,B=z+J,X=R-H,Z=z-J,U=k+Q,$=P+V,O=h*(k-Q),et=h*(P-V),ot=W+U,at=B+$,it=X+et,ct=Z-O;if(t[p]=ot,t[p+1]=at,t[v]=it,t[v+1]=ct,f===0){var lt=W-U,ht=B-$;t[w]=lt,t[w+1]=ht;continue}if(f!==g){var dt=X,It=-Z,Dt=W,jt=-B,Wt=-h*et,Bt=-h*O,Ut=-h*$,$t=-h*U,Pt=dt+Wt,Lt=It+Bt,Gt=Dt+$t,Ht=jt-Ut,yt=r+_-f,bt=r+m-f;t[yt]=Pt,t[yt+1]=Lt,t[bt]=Gt,t[bt+1]=Ht}}}},i.prototype._singleRealTransform2=function(t,e,o){const s=this._out,a=this._data,r=a[e],c=a[e+o],l=r+c,h=r-c;s[t]=l,s[t+1]=0,s[t+2]=h,s[t+3]=0},i.prototype._singleRealTransform4=function(t,e,o){const s=this._out,a=this._data,r=this._inv?-1:1,c=o*2,l=o*3,h=a[e],d=a[e+o],m=a[e+c],_=a[e+l],g=h+m,f=h-m,u=d+_,p=r*(d-_),v=g+u,w=f,b=-p,A=g-u,E=f,y=p;s[t]=v,s[t+1]=0,s[t+2]=w,s[t+3]=b,s[t+4]=A,s[t+5]=0,s[t+6]=E,s[t+7]=y},mt}var Ot=Zt();const qt=Xt(Ot),nt=16e3,j=512,D=64,At=Math.floor(.025*nt),Ft=Math.floor(.01*nt);function Mt(i){return 2595*Math.log10(1+i/700)}function te(i){return 700*(10**(i/2595)-1)}function ee(){const i=Mt(0),n=Mt(nt/2),t=new Float64Array(D+2);for(let r=0;r<D+2;r++)t[r]=i+r*(n-i)/(D+1);const o=t.map(r=>te(r)).map(r=>Math.floor((j+1)*r/nt)),s=[],a=Math.floor(j/2)+1;for(let r=0;r<D;r++){const c=new Float32Array(a);for(let l=o[r];l<o[r+1];l++)c[l]=(l-o[r])/(o[r+1]-o[r]);for(let l=o[r+1];l<o[r+2];l++)c[l]=(o[r+2]-l)/(o[r+2]-o[r+1]);s.push(c)}return s}const se=ee(),rt=new qt(j),ft=new Float32Array(j),Ct=rt.createComplexArray(),_t=rt.createComplexArray(),Tt=new Float32Array(Math.floor(j/2)+1);function Nt(i){const n=1+Math.ceil((i.length-At)/Ft),t=new Float32Array(n*D),e=Math.floor(j/2)+1;for(let o=0;o<n;o++){const s=o*Ft;ft.fill(0);for(let a=0;a<At&&s+a<i.length;a++)ft[a]=i[s+a];rt.toComplexArray(ft,Ct),rt.transform(_t,Ct);for(let a=0;a<e;a++){const r=_t[2*a],c=_t[2*a+1],l=(r*r+c*c)/j;Tt[a]=l===0?1e-30:l}for(let a=0;a<D;a++){const r=se[a];let c=0;for(let l=0;l<e;l++)c+=Tt[l]*r[l];t[o*D+a]=Math.log(c===0?1e-30:c)}}return t}function ne(i,n){let t=0;for(let e=0;e<i.length;e++)t+=i[e]*n[e];return(t+1)/2}function re(i,n){let t=0;for(const e of n){const o=ne(i,e);o>t&&(t=o)}return t}class xt extends EventTarget{constructor({name:n,refEmbeddings:t,threshold:e=.65,relaxationMs:o=2e3,inferenceGapMs:s=300}){super(),this.name=n,this.refEmbeddings=t,this.threshold=e,this.relaxationMs=o,this.inferenceGapMs=s,this._lastDetectionAt=0,this._lastInferenceAt=0,this._lastScore=0}get lastScore(){return this._lastScore}async scoreFrame(n){const t=Date.now();if(t-this._lastInferenceAt<this.inferenceGapMs)return null;this._lastInferenceAt=t;const e=Nt(n),o=await St(e),s=re(o,this.refEmbeddings);return this._lastScore=s,s>=this.threshold&&t-this._lastDetectionAt>=this.relaxationMs&&(this._lastDetectionAt=t,this.dispatchEvent(new CustomEvent("match",{detail:{name:this.name,confidence:s,timestamp:t}}))),s}}const Rt=16e3,oe=1500,pt=24e3;function zt(i){if(i.length===pt)return i;const n=new Float32Array(pt);return n.set(i.subarray(0,pt)),n}class kt extends EventTarget{constructor(n){super(),this.wordName=n.trim().toLowerCase(),this.samples=[]}get sampleCount(){return this.samples.length}async recordSample(){const n=await navigator.mediaDevices.getUserMedia({audio:!0});return new Promise((t,e)=>{const o=new AudioContext({sampleRate:Rt}),s=new MediaRecorder(n),a=[];this.dispatchEvent(new CustomEvent("recording-start")),s.ondataavailable=r=>{r.data.size>0&&a.push(r.data)},s.onstop=async()=>{n.getTracks().forEach(r=>r.stop());try{const c=await new Blob(a,{type:"audio/webm"}).arrayBuffer(),l=await o.decodeAudioData(c);await o.close();const h=l.getChannelData(0),d=zt(new Float32Array(h)),m=this._push(d,`Recorded #${this.samples.length}`);t(m)}catch(r){await o.close().catch(()=>{}),e(r)}},s.start(),setTimeout(()=>s.stop(),oe)})}async addAudioFile(n){const t=await n.arrayBuffer(),e=new AudioContext({sampleRate:Rt}),o=await e.decodeAudioData(t);await e.close();const s=o.getChannelData(0),a=zt(new Float32Array(s));return this._push(a,n.name)}removeSample(n){this.samples.splice(n,1),this.dispatchEvent(new CustomEvent("samples-changed",{detail:{count:this.samples.length}}))}clearSamples(){this.samples=[],this.dispatchEvent(new CustomEvent("samples-changed",{detail:{count:0}}))}async generateRef(){if(this.samples.length<3)throw new Error(`Need at least 3 samples (currently have ${this.samples.length})`);this.dispatchEvent(new CustomEvent("generating",{detail:{total:this.samples.length}}));const n=[];for(let t=0;t<this.samples.length;t++){const e=Nt(this.samples[t].audioBuffer),o=await St(e);n.push(Array.from(o)),this.dispatchEvent(new CustomEvent("progress",{detail:{done:t+1,total:this.samples.length}}))}return{word_name:this.wordName,model_type:"resnet_50_arc",embeddings:n}}_push(n,t){this.samples.push({audioBuffer:n,name:t});const e=this.samples.length;return this.dispatchEvent(new CustomEvent("sample-added",{detail:{count:e,name:t}})),e}}const ae=`/**
+ * public/audio-processor.js
+ * AudioWorklet that runs at 16 kHz and continuously emits the last
+ * 1.5-second window (24 000 samples) via a circular buffer.
+ *
+ * The main thread receives a fresh Float32Array on every
+ * AudioWorklet quantum (128 samples ≈ every 8 ms at 16 kHz).
+ * The inference loop in engine.js rate-limits to avoid excessive work.
+ */
+class AudioProcessor extends AudioWorkletProcessor {
+    constructor() {
+        super()
+        this._size = 24000  // 1.5 s × 16 000 Hz
+        this._buf = new Float32Array(this._size)
+        this._ptr = 0
+    }
+
+    process(inputs) {
+        const ch = inputs[0]?.[0]
+        if (!ch) return true
+
+        for (let i = 0; i < ch.length; i++) {
+            this._buf[this._ptr] = ch[i]
+            this._ptr = (this._ptr + 1) % this._size
+        }
+
+        // Send an ordered copy of the ring buffer
+        const out = new Float32Array(this._size)
+        for (let i = 0; i < this._size; i++) {
+            out[i] = this._buf[(this._ptr + i) % this._size]
+        }
+        this.port.postMessage(out)
+        return true
+    }
+}
+
+registerProcessor('audio-processor', AudioProcessor)
+`;let vt=null;function ie(){if(!vt){const i=new Blob([ae],{type:"application/javascript"});vt=URL.createObjectURL(i)}return vt}const wt="mellon_custom_refs";function gt(){try{const i=localStorage.getItem(wt);return i?JSON.parse(i):[]}catch{return[]}}function ce(i){const n=gt().filter(t=>t.word_name!==i.word_name);n.push(i),localStorage.setItem(wt,JSON.stringify(n))}function le(i){const n=gt().filter(t=>t.word_name!==i);localStorage.setItem(wt,JSON.stringify(n))}function he(i){const n=JSON.stringify(i,null,2),t=new Blob([n],{type:"application/json"}),e=URL.createObjectURL(t),o=Object.assign(document.createElement("a"),{href:e,download:`${i.word_name}_ref.json`});document.body.appendChild(o),o.click(),document.body.removeChild(o),URL.revokeObjectURL(e)}async function de(i){const n=await i.text();let t;try{t=JSON.parse(n)}catch{throw new Error("Invalid JSON")}if(!t.embeddings||!Array.isArray(t.embeddings)||!t.embeddings.length)throw new Error('Missing or empty "embeddings" array');if(!Array.isArray(t.embeddings[0]))throw new Error('"embeddings" must be a 2D array');return t.word_name||(t.word_name=i.name.replace(/_ref\.json$/i,"").replace(/\.json$/i,"")),t}class ue extends EventTarget{constructor(n={}){super(),this._opts={words:n.words??[],refs:n.refs??[],threshold:n.threshold??.65,relaxationMs:n.relaxationMs??2e3,inferenceGapMs:n.inferenceGapMs??300,assetsPath:n.assetsPath},this._refs=new Map,this._detectors=new Map,this._audioCtx=null,this._workletNode=null,this._stream=null,this._initialized=!1,this._running=!1}get isInitialized(){return this._initialized}get isRunning(){return this._running}async init(n){if(this._initialized){n==null||n(1);return}this._opts.assetsPath&&Qt({assetsPath:this._opts.assetsPath});try{await Vt(n)}catch(t){throw this.dispatchEvent(new CustomEvent("error",{detail:{error:t}})),t}for(const t of this._opts.refs)try{let e;if(typeof t=="string"){const o=await fetch(t);if(!o.ok)throw new Error(`HTTP ${o.status}`);e=await o.json()}else e=t;this.addCustomWord(e)}catch(e){const o=typeof t=="string"?t:t.word_name;console.warn(`[Mellon] Failed to load ref "${o}": ${e.message}`)}this._initialized=!0,this.dispatchEvent(new CustomEvent("ready"))}async start(n){this._initialized||await this.init();const t=n??this._opts.words;try{this._stream=await navigator.mediaDevices.getUserMedia({audio:!0})}catch(s){const a=new Error(`Microphone access denied: ${s.message}`);throw this.dispatchEvent(new CustomEvent("error",{detail:{error:a}})),a}this._audioCtx=new AudioContext({sampleRate:16e3});const e=ie();await this._audioCtx.audioWorklet.addModule(e);const o=this._audioCtx.createMediaStreamSource(this._stream);this._workletNode=new AudioWorkletNode(this._audioCtx,"audio-processor"),o.connect(this._workletNode),this._workletNode.connect(this._audioCtx.destination);for(const s of t){const a=this._refs.get(s);if(!a){console.warn(`[Mellon] No reference embeddings for "${s}" — skipping. Call addCustomWord() to register custom words before start().`);continue}const r=new xt({name:s,refEmbeddings:a.embeddings,threshold:this._opts.threshold,relaxationMs:this._opts.relaxationMs,inferenceGapMs:this._opts.inferenceGapMs});r.addEventListener("match",c=>{this.dispatchEvent(new CustomEvent("match",{detail:c.detail}))}),this._detectors.set(s,r)}this._workletNode.port.onmessage=async s=>{const a=[];for(const r of this._detectors.values())a.push(r.scoreFrame(s.data));await Promise.allSettled(a)},this._running=!0}stop(){this._workletNode&&(this._workletNode.port.onmessage=null,this._workletNode.disconnect(),this._workletNode=null),this._stream&&(this._stream.getTracks().forEach(n=>n.stop()),this._stream=null),this._audioCtx&&(this._audioCtx.close(),this._audioCtx=null),this._detectors.clear(),this._running=!1}addCustomWord(n){if(this._refs.set(n.word_name,n),this._running&&this._workletNode){const t=new xt({name:n.word_name,refEmbeddings:n.embeddings,threshold:this._opts.threshold,relaxationMs:this._opts.relaxationMs,inferenceGapMs:this._opts.inferenceGapMs});t.addEventListener("match",e=>{this.dispatchEvent(new CustomEvent("match",{detail:e.detail}))}),this._detectors.set(n.word_name,t)}}enrollWord(n){return new kt(n)}static loadWords(){return gt()}static saveWord(n){ce(n)}static deleteWord(n){le(n)}static importWordFile(n){return de(n)}static exportWord(n){he(n)}}exports.EnrollmentSession=kt;exports.Mellon=ue;

package/dist/mellon.mjs

@@ -0,0 +1,611 @@
+const Ht = "0.0.5", Jt = [1, 1, 149, 64], Kt = `https://cdn.jsdelivr.net/npm/mellon@${Ht}/dist/assets`, st = {
+  assetsPath: `${Kt}`
+};
+let I = null, q = null, tt = null;
+function Yt({ assetsPath: i } = {}) {
+  i !== void 0 && (st.assetsPath = i), I = null, q = null, tt = null;
+}
+async function Qt(i) {
+  return I ? (i == null || i(1), I) : q || (q = (async () => {
+    const n = st.assetsPath.endsWith("/") ? st.assetsPath : st.assetsPath + "/", t = n + "ort.all.min.mjs", e = n + "model.onnx";
+    tt = await new Function("url", "return import(url)")(t), tt.env.wasm.wasmPaths = n;
+    const s = await fetch(e);
+    if (!s.ok) throw new Error(`Failed to fetch model: ${s.status}`);
+    const a = parseInt(s.headers.get("content-length") || "0", 10), r = s.body.getReader(), c = [];
+    let l = 0;
+    for (; ; ) {
+      const { done: m, value: _ } = await r.read();
+      if (m) break;
+      c.push(_), l += _.byteLength, a > 0 && (i == null || i(l / a));
+    }
+    const h = new Uint8Array(l);
+    let d = 0;
+    for (const m of c)
+      h.set(m, d), d += m.byteLength;
+    return I = await tt.InferenceSession.create(h.buffer, {
+      executionProviders: ["wasm"],
+      graphOptimizationLevel: "all"
+    }), i == null || i(1), I;
+  })(), q);
+}
+async function St(i) {
+  if (!I) throw new Error("Model not loaded — call loadModel() first");
+  const n = new tt.Tensor("float32", i, Jt), t = await I.run({ input: n }), e = Object.keys(t)[0];
+  return t[e].data;
+}
+function Vt(i) {
+  return i && i.__esModule && Object.prototype.hasOwnProperty.call(i, "default") ? i.default : i;
+}
+var mt, Et;
+function Xt() {
+  if (Et) return mt;
+  Et = 1;
+  function i(n) {
+    if (this.size = n | 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 = n << 1;
+    for (var t = new Array(this.size * 2), e = 0; e < t.length; e += 2) {
+      const l = Math.PI * e / this.size;
+      t[e] = Math.cos(l), t[e + 1] = -Math.sin(l);
+    }
+    this.table = t;
+    for (var o = 0, s = 1; this.size > s; s <<= 1)
+      o++;
+    this._width = o % 2 === 0 ? o - 1 : o, this._bitrev = new Array(1 << this._width);
+    for (var a = 0; a < this._bitrev.length; a++) {
+      this._bitrev[a] = 0;
+      for (var r = 0; r < this._width; r += 2) {
+        var c = this._width - r - 2;
+        this._bitrev[a] |= (a >>> r & 3) << c;
+      }
+    }
+    this._out = null, this._data = null, this._inv = 0;
+  }
+  return mt = i, i.prototype.fromComplexArray = function(t, e) {
+    for (var o = e || new Array(t.length >>> 1), s = 0; s < t.length; s += 2)
+      o[s >>> 1] = t[s];
+    return o;
+  }, i.prototype.createComplexArray = function() {
+    const t = new Array(this._csize);
+    for (var e = 0; e < t.length; e++)
+      t[e] = 0;
+    return t;
+  }, i.prototype.toComplexArray = function(t, e) {
+    for (var o = e || this.createComplexArray(), s = 0; s < o.length; s += 2)
+      o[s] = t[s >>> 1], o[s + 1] = 0;
+    return o;
+  }, i.prototype.completeSpectrum = function(t) {
+    for (var e = this._csize, o = e >>> 1, s = 2; s < o; s += 2)
+      t[e - s] = t[s], t[e - s + 1] = -t[s + 1];
+  }, i.prototype.transform = function(t, e) {
+    if (t === e)
+      throw new Error("Input and output buffers must be different");
+    this._out = t, this._data = e, this._inv = 0, this._transform4(), this._out = null, this._data = null;
+  }, i.prototype.realTransform = function(t, e) {
+    if (t === e)
+      throw new Error("Input and output buffers must be different");
+    this._out = t, this._data = e, this._inv = 0, this._realTransform4(), this._out = null, this._data = null;
+  }, i.prototype.inverseTransform = function(t, e) {
+    if (t === e)
+      throw new Error("Input and output buffers must be different");
+    this._out = t, this._data = e, this._inv = 1, this._transform4();
+    for (var o = 0; o < t.length; o++)
+      t[o] /= this.size;
+    this._out = null, this._data = null;
+  }, i.prototype._transform4 = function() {
+    var t = this._out, e = this._csize, o = this._width, s = 1 << o, a = e / s << 1, r, c, l = this._bitrev;
+    if (a === 4)
+      for (r = 0, c = 0; r < e; r += a, c++) {
+        const u = l[c];
+        this._singleTransform2(r, u, s);
+      }
+    else
+      for (r = 0, c = 0; r < e; r += a, c++) {
+        const u = l[c];
+        this._singleTransform4(r, u, s);
+      }
+    var h = this._inv ? -1 : 1, d = this.table;
+    for (s >>= 2; s >= 2; s >>= 2) {
+      a = e / s << 1;
+      var m = a >>> 2;
+      for (r = 0; r < e; r += a)
+        for (var _ = r + m, g = r, f = 0; g < _; g += 2, f += s) {
+          const u = g, p = u + m, v = p + m, w = v + m, b = t[u], A = t[u + 1], E = t[p], y = t[p + 1], F = t[v], C = t[v + 1], M = t[w], T = t[w + 1], x = b, R = A, z = d[f], S = h * d[f + 1], N = E * z - y * S, k = E * S + y * z, L = d[2 * f], P = h * d[2 * f + 1], G = F * L - C * P, H = F * P + C * L, J = d[3 * f], K = h * d[3 * f + 1], Y = M * J - T * K, Q = M * K + T * J, V = x + G, j = R + H, B = x - G, X = R - H, Z = N + Y, U = k + Q, $ = h * (N - Y), O = h * (k - Q), et = V + Z, ot = j + U, at = V - Z, it = j - U, ct = B + O, lt = X - $, ht = B - O, dt = X + $;
+          t[u] = et, t[u + 1] = ot, t[p] = ct, t[p + 1] = lt, t[v] = at, t[v + 1] = it, t[w] = ht, t[w + 1] = dt;
+        }
+    }
+  }, i.prototype._singleTransform2 = function(t, e, o) {
+    const s = this._out, a = this._data, r = a[e], c = a[e + 1], l = a[e + o], h = a[e + o + 1], d = r + l, m = c + h, _ = r - l, g = c - h;
+    s[t] = d, s[t + 1] = m, s[t + 2] = _, s[t + 3] = g;
+  }, i.prototype._singleTransform4 = function(t, e, o) {
+    const s = this._out, a = this._data, r = this._inv ? -1 : 1, c = o * 2, l = o * 3, h = a[e], d = a[e + 1], m = a[e + o], _ = a[e + o + 1], g = a[e + c], f = a[e + c + 1], u = a[e + l], p = a[e + l + 1], v = h + g, w = d + f, b = h - g, A = d - f, E = m + u, y = _ + p, F = r * (m - u), C = r * (_ - p), M = v + E, T = w + y, x = b + C, R = A - F, z = v - E, S = w - y, N = b - C, k = A + F;
+    s[t] = M, s[t + 1] = T, s[t + 2] = x, s[t + 3] = R, s[t + 4] = z, s[t + 5] = S, s[t + 6] = N, s[t + 7] = k;
+  }, i.prototype._realTransform4 = function() {
+    var t = this._out, e = this._csize, o = this._width, s = 1 << o, a = e / s << 1, r, c, l = this._bitrev;
+    if (a === 4)
+      for (r = 0, c = 0; r < e; r += a, c++) {
+        const ut = l[c];
+        this._singleRealTransform2(r, ut >>> 1, s >>> 1);
+      }
+    else
+      for (r = 0, c = 0; r < e; r += a, c++) {
+        const ut = l[c];
+        this._singleRealTransform4(r, ut >>> 1, s >>> 1);
+      }
+    var h = this._inv ? -1 : 1, d = this.table;
+    for (s >>= 2; s >= 2; s >>= 2) {
+      a = e / s << 1;
+      var m = a >>> 1, _ = m >>> 1, g = _ >>> 1;
+      for (r = 0; r < e; r += a)
+        for (var f = 0, u = 0; f <= g; f += 2, u += s) {
+          var p = r + f, v = p + _, w = v + _, b = w + _, A = t[p], E = t[p + 1], y = t[v], F = t[v + 1], C = t[w], M = t[w + 1], T = t[b], x = t[b + 1], R = A, z = E, S = d[u], N = h * d[u + 1], k = y * S - F * N, L = y * N + F * S, P = d[2 * u], G = h * d[2 * u + 1], H = C * P - M * G, J = C * G + M * P, K = d[3 * u], Y = h * d[3 * u + 1], Q = T * K - x * Y, V = T * Y + x * K, j = R + H, B = z + J, X = R - H, Z = z - J, U = k + Q, $ = L + V, O = h * (k - Q), et = h * (L - V), ot = j + U, at = B + $, it = X + et, ct = Z - O;
+          if (t[p] = ot, t[p + 1] = at, t[v] = it, t[v + 1] = ct, f === 0) {
+            var lt = j - U, ht = B - $;
+            t[w] = lt, t[w + 1] = ht;
+            continue;
+          }
+          if (f !== g) {
+            var dt = X, kt = -Z, It = j, Dt = -B, Wt = -h * et, jt = -h * O, Bt = -h * $, Ut = -h * U, $t = dt + Wt, Lt = kt + jt, Pt = It + Ut, Gt = Dt - Bt, yt = r + _ - f, bt = r + m - f;
+            t[yt] = $t, t[yt + 1] = Lt, t[bt] = Pt, t[bt + 1] = Gt;
+          }
+        }
+    }
+  }, i.prototype._singleRealTransform2 = function(t, e, o) {
+    const s = this._out, a = this._data, r = a[e], c = a[e + o], l = r + c, h = r - c;
+    s[t] = l, s[t + 1] = 0, s[t + 2] = h, s[t + 3] = 0;
+  }, i.prototype._singleRealTransform4 = function(t, e, o) {
+    const s = this._out, a = this._data, r = this._inv ? -1 : 1, c = o * 2, l = o * 3, h = a[e], d = a[e + o], m = a[e + c], _ = a[e + l], g = h + m, f = h - m, u = d + _, p = r * (d - _), v = g + u, w = f, b = -p, A = g - u, E = f, y = p;
+    s[t] = v, s[t + 1] = 0, s[t + 2] = w, s[t + 3] = b, s[t + 4] = A, s[t + 5] = 0, s[t + 6] = E, s[t + 7] = y;
+  }, mt;
+}
+var Zt = Xt();
+const Ot = /* @__PURE__ */ Vt(Zt), nt = 16e3, W = 512, D = 64, At = Math.floor(0.025 * nt), Ft = Math.floor(0.01 * nt);
+function Ct(i) {
+  return 2595 * Math.log10(1 + i / 700);
+}
+function qt(i) {
+  return 700 * (10 ** (i / 2595) - 1);
+}
+function te() {
+  const i = Ct(0), n = Ct(nt / 2), t = new Float64Array(D + 2);
+  for (let r = 0; r < D + 2; r++)
+    t[r] = i + r * (n - i) / (D + 1);
+  const o = t.map((r) => qt(r)).map((r) => Math.floor((W + 1) * r / nt)), s = [], a = Math.floor(W / 2) + 1;
+  for (let r = 0; r < D; r++) {
+    const c = new Float32Array(a);
+    for (let l = o[r]; l < o[r + 1]; l++) c[l] = (l - o[r]) / (o[r + 1] - o[r]);
+    for (let l = o[r + 1]; l < o[r + 2]; l++) c[l] = (o[r + 2] - l) / (o[r + 2] - o[r + 1]);
+    s.push(c);
+  }
+  return s;
+}
+const ee = te(), rt = new Ot(W), ft = new Float32Array(W), Mt = rt.createComplexArray(), _t = rt.createComplexArray(), Tt = new Float32Array(Math.floor(W / 2) + 1);
+function Nt(i) {
+  const n = 1 + Math.ceil((i.length - At) / Ft), t = new Float32Array(n * D), e = Math.floor(W / 2) + 1;
+  for (let o = 0; o < n; o++) {
+    const s = o * Ft;
+    ft.fill(0);
+    for (let a = 0; a < At && s + a < i.length; a++)
+      ft[a] = i[s + a];
+    rt.toComplexArray(ft, Mt), rt.transform(_t, Mt);
+    for (let a = 0; a < e; a++) {
+      const r = _t[2 * a], c = _t[2 * a + 1], l = (r * r + c * c) / W;
+      Tt[a] = l === 0 ? 1e-30 : l;
+    }
+    for (let a = 0; a < D; a++) {
+      const r = ee[a];
+      let c = 0;
+      for (let l = 0; l < e; l++) c += Tt[l] * r[l];
+      t[o * D + a] = Math.log(c === 0 ? 1e-30 : c);
+    }
+  }
+  return t;
+}
+function se(i, n) {
+  let t = 0;
+  for (let e = 0; e < i.length; e++) t += i[e] * n[e];
+  return (t + 1) / 2;
+}
+function ne(i, n) {
+  let t = 0;
+  for (const e of n) {
+    const o = se(i, e);
+    o > t && (t = o);
+  }
+  return t;
+}
+class xt extends EventTarget {
+  /**
+   * @param {object} opts
+   * @param {string}              opts.name           Human-readable label
+   * @param {number[][]|Float32Array[]} opts.refEmbeddings Reference embeddings
+   * @param {number}              [opts.threshold=0.65]  Detection threshold
+   * @param {number}              [opts.relaxationMs=2000] Min ms between events
+   * @param {number}              [opts.inferenceGapMs=300] Min ms between inferences
+   */
+  constructor({ name: n, refEmbeddings: t, threshold: e = 0.65, relaxationMs: o = 2e3, inferenceGapMs: s = 300 }) {
+    super(), this.name = n, this.refEmbeddings = t, this.threshold = e, this.relaxationMs = o, this.inferenceGapMs = s, this._lastDetectionAt = 0, this._lastInferenceAt = 0, this._lastScore = 0;
+  }
+  get lastScore() {
+    return this._lastScore;
+  }
+  /**
+   * Score one 1.5-second audio frame. Rate-limited to `inferenceGapMs`.
+   *
+   * @param {Float32Array} audioBuffer  24 000 samples at 16 kHz
+   * @returns {Promise<number|null>}    Similarity score, or null if rate-limited
+   */
+  async scoreFrame(n) {
+    const t = Date.now();
+    if (t - this._lastInferenceAt < this.inferenceGapMs) return null;
+    this._lastInferenceAt = t;
+    const e = Nt(n), o = await St(e), s = ne(o, this.refEmbeddings);
+    return this._lastScore = s, s >= this.threshold && t - this._lastDetectionAt >= this.relaxationMs && (this._lastDetectionAt = t, this.dispatchEvent(new CustomEvent("match", {
+      detail: { name: this.name, confidence: s, timestamp: t }
+    }))), s;
+  }
+}
+const Rt = 16e3, re = 1500, pt = 24e3;
+function zt(i) {
+  if (i.length === pt) return i;
+  const n = new Float32Array(pt);
+  return n.set(i.subarray(0, pt)), n;
+}
+class oe extends EventTarget {
+  /** @param {string} wordName — the wake word label */
+  constructor(n) {
+    super(), this.wordName = n.trim().toLowerCase(), this.samples = [];
+  }
+  get sampleCount() {
+    return this.samples.length;
+  }
+  // ─── Recording ─────────────────────────────────────────────────────────────
+  /**
+   * Record exactly `RECORD_MS` milliseconds from the microphone.
+   * Dispatches 'recording-start' and 'sample-added' events.
+   *
+   * @returns {Promise<number>}  Index (1-based) of the new sample
+   */
+  async recordSample() {
+    const n = await navigator.mediaDevices.getUserMedia({ audio: !0 });
+    return new Promise((t, e) => {
+      const o = new AudioContext({ sampleRate: Rt }), s = new MediaRecorder(n), a = [];
+      this.dispatchEvent(new CustomEvent("recording-start")), s.ondataavailable = (r) => {
+        r.data.size > 0 && a.push(r.data);
+      }, s.onstop = async () => {
+        n.getTracks().forEach((r) => r.stop());
+        try {
+          const c = await new Blob(a, { type: "audio/webm" }).arrayBuffer(), l = await o.decodeAudioData(c);
+          await o.close();
+          const h = l.getChannelData(0), d = zt(new Float32Array(h)), m = this._push(d, `Recorded #${this.samples.length}`);
+          t(m);
+        } catch (r) {
+          await o.close().catch(() => {
+          }), e(r);
+        }
+      }, s.start(), setTimeout(() => s.stop(), re);
+    });
+  }
+  // ─── Upload ────────────────────────────────────────────────────────────────
+  /**
+   * Decode an uploaded audio File and add it as a sample.
+   *
+   * @param {File} file
+   * @returns {Promise<number>}  Index (1-based) of the new sample
+   */
+  async addAudioFile(n) {
+    const t = await n.arrayBuffer(), e = new AudioContext({ sampleRate: Rt }), o = await e.decodeAudioData(t);
+    await e.close();
+    const s = o.getChannelData(0), a = zt(new Float32Array(s));
+    return this._push(a, n.name);
+  }
+  // ─── Manage ────────────────────────────────────────────────────────────────
+  /**
+   * Remove a sample by 0-based index.
+   * @param {number} idx
+   */
+  removeSample(n) {
+    this.samples.splice(n, 1), this.dispatchEvent(new CustomEvent("samples-changed", { detail: { count: this.samples.length } }));
+  }
+  clearSamples() {
+    this.samples = [], this.dispatchEvent(new CustomEvent("samples-changed", { detail: { count: 0 } }));
+  }
+  // ─── Generate ──────────────────────────────────────────────────────────────
+  /**
+   * Compute embeddings for all collected samples and return a ref object.
+   * Dispatches 'progress' events during embedding.
+   *
+   * @returns {Promise<{ word_name:string, model_type:string, embeddings:number[][] }>}
+   */
+  async generateRef() {
+    if (this.samples.length < 3)
+      throw new Error(`Need at least 3 samples (currently have ${this.samples.length})`);
+    this.dispatchEvent(new CustomEvent("generating", { detail: { total: this.samples.length } }));
+    const n = [];
+    for (let t = 0; t < this.samples.length; t++) {
+      const e = Nt(this.samples[t].audioBuffer), o = await St(e);
+      n.push(Array.from(o)), this.dispatchEvent(new CustomEvent("progress", {
+        detail: { done: t + 1, total: this.samples.length }
+      }));
+    }
+    return {
+      word_name: this.wordName,
+      model_type: "resnet_50_arc",
+      embeddings: n
+    };
+  }
+  // ─── Private ───────────────────────────────────────────────────────────────
+  _push(n, t) {
+    this.samples.push({ audioBuffer: n, name: t });
+    const e = this.samples.length;
+    return this.dispatchEvent(new CustomEvent("sample-added", { detail: { count: e, name: t } })), e;
+  }
+}
+const ae = `/**
+ * public/audio-processor.js
+ * AudioWorklet that runs at 16 kHz and continuously emits the last
+ * 1.5-second window (24 000 samples) via a circular buffer.
+ *
+ * The main thread receives a fresh Float32Array on every
+ * AudioWorklet quantum (128 samples ≈ every 8 ms at 16 kHz).
+ * The inference loop in engine.js rate-limits to avoid excessive work.
+ */
+class AudioProcessor extends AudioWorkletProcessor {
+    constructor() {
+        super()
+        this._size = 24000  // 1.5 s × 16 000 Hz
+        this._buf = new Float32Array(this._size)
+        this._ptr = 0
+    }
+
+    process(inputs) {
+        const ch = inputs[0]?.[0]
+        if (!ch) return true
+
+        for (let i = 0; i < ch.length; i++) {
+            this._buf[this._ptr] = ch[i]
+            this._ptr = (this._ptr + 1) % this._size
+        }
+
+        // Send an ordered copy of the ring buffer
+        const out = new Float32Array(this._size)
+        for (let i = 0; i < this._size; i++) {
+            out[i] = this._buf[(this._ptr + i) % this._size]
+        }
+        this.port.postMessage(out)
+        return true
+    }
+}
+
+registerProcessor('audio-processor', AudioProcessor)
+`;
+let vt = null;
+function ie() {
+  if (!vt) {
+    const i = new Blob([ae], { type: "application/javascript" });
+    vt = URL.createObjectURL(i);
+  }
+  return vt;
+}
+const wt = "mellon_custom_refs";
+function gt() {
+  try {
+    const i = localStorage.getItem(wt);
+    return i ? JSON.parse(i) : [];
+  } catch {
+    return [];
+  }
+}
+function ce(i) {
+  const n = gt().filter((t) => t.word_name !== i.word_name);
+  n.push(i), localStorage.setItem(wt, JSON.stringify(n));
+}
+function le(i) {
+  const n = gt().filter((t) => t.word_name !== i);
+  localStorage.setItem(wt, JSON.stringify(n));
+}
+function he(i) {
+  const n = JSON.stringify(i, null, 2), t = new Blob([n], { type: "application/json" }), e = URL.createObjectURL(t), o = Object.assign(document.createElement("a"), {
+    href: e,
+    download: `${i.word_name}_ref.json`
+  });
+  document.body.appendChild(o), o.click(), document.body.removeChild(o), URL.revokeObjectURL(e);
+}
+async function de(i) {
+  const n = await i.text();
+  let t;
+  try {
+    t = JSON.parse(n);
+  } catch {
+    throw new Error("Invalid JSON");
+  }
+  if (!t.embeddings || !Array.isArray(t.embeddings) || !t.embeddings.length)
+    throw new Error('Missing or empty "embeddings" array');
+  if (!Array.isArray(t.embeddings[0]))
+    throw new Error('"embeddings" must be a 2D array');
+  return t.word_name || (t.word_name = i.name.replace(/_ref\.json$/i, "").replace(/\.json$/i, "")), t;
+}
+class ue extends EventTarget {
+  /**
+   * @param {object}   [opts]
+   * @param {string[]} [opts.words]              Words to detect (must have refs loaded via addCustomWord())
+   * @param {Array<string|{word_name:string,embeddings:number[][]}>} [opts.refs]
+   *   Reference data to preload. Each entry is either a URL string pointing to a
+   *   hosted `_ref.json` file, or an inline RefData object.
+   *   Refs are fetched/loaded during init() before detection starts.
+   * @param {number}   [opts.threshold=0.65]     Detection threshold [0, 1]
+   * @param {number}   [opts.relaxationMs=2000]  Min ms between successive match events
+   * @param {number}   [opts.inferenceGapMs=300] Min ms between inference runs
+   * @param {string}   [opts.assetsPath]  
+   */
+  constructor(n = {}) {
+    super(), this._opts = {
+      words: n.words ?? [],
+      refs: n.refs ?? [],
+      threshold: n.threshold ?? 0.65,
+      relaxationMs: n.relaxationMs ?? 2e3,
+      inferenceGapMs: n.inferenceGapMs ?? 300,
+      assetsPath: n.assetsPath
+    }, this._refs = /* @__PURE__ */ new Map(), this._detectors = /* @__PURE__ */ new Map(), this._audioCtx = null, this._workletNode = null, this._stream = null, this._initialized = !1, this._running = !1;
+  }
+  /** Whether init() has completed successfully. */
+  get isInitialized() {
+    return this._initialized;
+  }
+  /** Whether start() is active (mic is open). */
+  get isRunning() {
+    return this._running;
+  }
+  // ─── Lifecycle ───────────────────────────────────────────────────────────
+  /**
+   * Load the ONNX model and pre-cache built-in reference embeddings.
+   * Calling this explicitly is optional — start() will auto-init if needed.
+   * Call it early to show a loading progress bar.
+   *
+   * @param {(progress: number) => void} [onProgress]  0.0 → 1.0
+   */
+  async init(n) {
+    if (this._initialized) {
+      n == null || n(1);
+      return;
+    }
+    this._opts.assetsPath && Yt({ assetsPath: this._opts.assetsPath });
+    try {
+      await Qt(n);
+    } catch (t) {
+      throw this.dispatchEvent(new CustomEvent("error", { detail: { error: t } })), t;
+    }
+    for (const t of this._opts.refs)
+      try {
+        let e;
+        if (typeof t == "string") {
+          const o = await fetch(t);
+          if (!o.ok) throw new Error(`HTTP ${o.status}`);
+          e = await o.json();
+        } else
+          e = t;
+        this.addCustomWord(e);
+      } catch (e) {
+        const o = typeof t == "string" ? t : t.word_name;
+        console.warn(`[Mellon] Failed to load ref "${o}": ${e.message}`);
+      }
+    this._initialized = !0, this.dispatchEvent(new CustomEvent("ready"));
+  }
+  /**
+   * Request microphone access and start hotword detection.
+   * Emits 'match' CustomEvents when a word is detected.
+   *
+   * @param {string[]} [words]  Subset of words to listen for; defaults to opts.words
+   */
+  async start(n) {
+    this._initialized || await this.init();
+    const t = n ?? this._opts.words;
+    try {
+      this._stream = await navigator.mediaDevices.getUserMedia({ audio: !0 });
+    } catch (s) {
+      const a = new Error(`Microphone access denied: ${s.message}`);
+      throw this.dispatchEvent(new CustomEvent("error", { detail: { error: a } })), a;
+    }
+    this._audioCtx = new AudioContext({ sampleRate: 16e3 });
+    const e = ie();
+    await this._audioCtx.audioWorklet.addModule(e);
+    const o = this._audioCtx.createMediaStreamSource(this._stream);
+    this._workletNode = new AudioWorkletNode(this._audioCtx, "audio-processor"), o.connect(this._workletNode), this._workletNode.connect(this._audioCtx.destination);
+    for (const s of t) {
+      const a = this._refs.get(s);
+      if (!a) {
+        console.warn(`[Mellon] No reference embeddings for "${s}" — skipping. Call addCustomWord() to register custom words before start().`);
+        continue;
+      }
+      const r = new xt({
+        name: s,
+        refEmbeddings: a.embeddings,
+        threshold: this._opts.threshold,
+        relaxationMs: this._opts.relaxationMs,
+        inferenceGapMs: this._opts.inferenceGapMs
+      });
+      r.addEventListener("match", (c) => {
+        this.dispatchEvent(new CustomEvent("match", { detail: c.detail }));
+      }), this._detectors.set(s, r);
+    }
+    this._workletNode.port.onmessage = async (s) => {
+      const a = [];
+      for (const r of this._detectors.values())
+        a.push(r.scoreFrame(s.data));
+      await Promise.allSettled(a);
+    }, this._running = !0;
+  }
+  /**
+   * Stop detection and release the microphone and AudioContext.
+   */
+  stop() {
+    this._workletNode && (this._workletNode.port.onmessage = null, this._workletNode.disconnect(), this._workletNode = null), this._stream && (this._stream.getTracks().forEach((n) => n.stop()), this._stream = null), this._audioCtx && (this._audioCtx.close(), this._audioCtx = null), this._detectors.clear(), this._running = !1;
+  }
+  // ─── Custom words ────────────────────────────────────────────────────────
+  /**
+   * Register reference embeddings for a custom (or overridden) word.
+   * Can be called before or after start(). If called while running, the new
+   * detector is added to the active pipeline without restarting.
+   *
+   * @param {{ word_name: string, model_type: string, embeddings: number[][] }} refData
+   */
+  addCustomWord(n) {
+    if (this._refs.set(n.word_name, n), this._running && this._workletNode) {
+      const t = new xt({
+        name: n.word_name,
+        refEmbeddings: n.embeddings,
+        threshold: this._opts.threshold,
+        relaxationMs: this._opts.relaxationMs,
+        inferenceGapMs: this._opts.inferenceGapMs
+      });
+      t.addEventListener("match", (e) => {
+        this.dispatchEvent(new CustomEvent("match", { detail: e.detail }));
+      }), this._detectors.set(n.word_name, t);
+    }
+  }
+  /**
+   * Create an EnrollmentSession for recording a new custom word.
+   * After generating the ref via session.generateRef(), pass the result
+   * to addCustomWord().
+   *
+   * @param {string} wordName
+   * @returns {EnrollmentSession}
+   *
+   * @example
+   *   const session = stt.enrollWord('hello')
+   *   await session.recordSample()   // record 1.5 s × 3+ times
+   *   const ref = await session.generateRef()
+   *   stt.addCustomWord(ref)
+   */
+  enrollWord(n) {
+    return new oe(n);
+  }
+  // ─── Persistence (static) ────────────────────────────────────────────────
+  /** Return all custom word refs stored in localStorage. */
+  static loadWords() {
+    return gt();
+  }
+  /** Persist a word ref to localStorage (replaces any existing entry with the same name). */
+  static saveWord(n) {
+    ce(n);
+  }
+  /** Delete a word ref from localStorage by name. */
+  static deleteWord(n) {
+    le(n);
+  }
+  /**
+   * Parse an uploaded ref JSON file and return a RefData object.
+   * @param {File} file
+   */
+  static importWordFile(n) {
+    return de(n);
+  }
+  /** Trigger a browser download of a ref as a JSON file. */
+  static exportWord(n) {
+    he(n);
+  }
+}
+export {
+  oe as EnrollmentSession,
+  ue as Mellon
+};

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mellon",
-  "version": "0.0.3",
+  "version": "0.0.5",
   "description": "Offline, in-browser hotword detection powered by EfficientWord-Net (ResNet-50 ArcFace). Works as a standalone app or npm library.",
   "type": "module",
   "main": "./dist/mellon.cjs",
@@ -17,8 +17,7 @@
     "dist/mellon.mjs",
     "dist/mellon.cjs",
     "dist/index.d.ts",
-    "dist/wasm",
-    "dist/models",
+    "dist/assets",
     "README.md"
   ],
   "keywords": [
@@ -33,13 +32,7 @@
     "voice"
   ],
   "scripts": {
-    "dev": "vite",
-    "build": "npm run build:app",
-    "build:app": "vite build && npm run copy-wasm",
-    "build:lib": "vite build --config vite.lib.config.js",
-    "copy-wasm": "node scripts/copy-wasm.js",
-    "preview": "vite preview",
-    "prepare": "node scripts/prepare_check.js"
+    "build": "vite build"
   },
   "dependencies": {
     "fft.js": "^4.0.4",