longpipe 0.0.14 → 0.0.16

package/README.md

@@ -1,286 +1,65 @@
 # Longpipe 🐉
 
-Fast, high quality virtual effects in the browser
+Fast, high-quality video effects in the browser — virtual backgrounds, face touch-up, auto-reframe, and mic noise removal in one open-source SDK.
 
+<img width="960" height="540" alt="video-effects-sdk" src="https://github.com/user-attachments/assets/5dbe1ccb-09dd-4f22-b4be-e3122767dbcc" />
 
-<img width="640" height="360" alt="0087" src="https://github.com/user-attachments/assets/525f983c-6b3d-43ff-aedf-b1203e2955c1" />
-
-
-
-Try the live demo: [longpipe.dev/demo](https://longpipe.dev/demo)
+**Try the live demo → [longpipe.dev/demo](https://longpipe.dev/demo)**
 
 > **Warning** — This project is very new and still under active development. Expect API changes between versions and bugs.
 
-## Features
-
-- **Support across browsers:** Works on every browser, simplifying dozens of browser inconsistencies in one simple API.
-- **Performance:** Longpipe was built from the ground up to work as well on 10-year-old netbooks as it does on the latest MacBook Pro.
-- **Adaptivity:** Longpipe has several model variants (xl to xs), auto-selecting to provide the best quality while maintaining 30 fps.
-- **Effects that compose:** virtual backgrounds, skin touch-up, and auto-reframe all run together off a single shared encoder pass — enabling one doesn't cost you another inference.
-- Built by the founder of [Vectorly](https://www.crunchbase.com/organization/vectorly), a commercial effects SDK acquired in 2021. A ground-up redesign for the WebGPU era, with years of production lessons baked in.
-
-## Install
+## Quick start
 
 ```
 npm install longpipe
 ```
 
-## Quick start
-
 ```ts
 import { EffectsPipeline } from 'longpipe'
 
-const inputStream = await navigator.mediaDevices.getUserMedia({ video: true, audio: true })
+const stream = await navigator.mediaDevices.getUserMedia({ video: true, audio: true })
 
-const pipeline = new EffectsPipeline(inputStream, {
-  background: 'blur',
-  audio: 'denoise'
+const pipeline = new EffectsPipeline(stream, {
+  background: 'blur',   // or an image, video, or solid color
+  touchup: true,        // skin smoothing
+  reframe: true,        // auto-frame the subject
+  audio: 'denoise',     // mic noise removal
 })
 
 videoEl.srcObject = pipeline.stream   // available immediately
 await pipeline.ready                  // optional — resolves once the effect is live
 ```
 
-`pipeline.stream` is wired synchronously and emits the unprocessed input until the model is ready (~1–3s on cold start, depending on hardware), so the consumer sees live video the whole time.
-
-
-## Performance
-
-Longpipe uses custom-trained models built with an EfficientNet-Lite encoder and a U-Net style decoder, broken into five performance presets (`xs` through `xl`) that vary in encoder size, decoder width, and input resolution. Across all variants Longpipe has much higher segmentation accuracy than alternative open-source models like MediaPipe and BodyPix, while also delivering much better real-world performance — the model runs as pure WebGPU/WebGL shaders in a zero-copy, fully-GPU pipeline.
-
-### Quality
-Using average alpha pixel error on the P3M-10K valdation dataset (499 landscape images), all variants of Longpipe surpassed mediapipe in both MAE and IoU.
-
-<img width="640" alt="lp-quality" src="https://github.com/user-attachments/assets/c7044937-95af-496f-893a-a05a50d7c914" />
-
-### Speed
-With the pure GPU zero copy pipeline, Longpipe achieves better real world performance than mediapipe using much larger models.
-
-<img width="640" alt="lp-speed" src="https://github.com/user-attachments/assets/ce658b8d-dc62-4d6b-90f8-cd62dfef2a59" />
-
-
-
-
-
-## Backgrounds
-
-The `background` option takes a wide set of inputs and normalizes them internally:
-
-```ts
-new EffectsPipeline(stream, { background: 'none' })                 // passthrough
-new EffectsPipeline(stream, { background: 'blur' })                 // default sigma
-new EffectsPipeline(stream, { background: { blur: { sigma: 12 } } })
-
-// images: URL, <img>, ImageBitmap, or { image: ... }
-new EffectsPipeline(stream, { background: 'https://example.com/bg.jpg' })
-new EffectsPipeline(stream, { background: imgElement })
-new EffectsPipeline(stream, { background: imageBitmap })
-
-// video: URL, Blob, or <video> — looped, muted, decoded on the main thread
-new EffectsPipeline(stream, { background: { video: 'https://example.com/bg.mp4' } })
-
-// solid color — hex string or [r, g, b] floats in [0, 1]
-new EffectsPipeline(stream, { background: { color: '#00b050' } })       // greenscreen
-new EffectsPipeline(stream, { background: { color: [0, 0.7, 0.3] } })
-```
-
-Swap at runtime:
+`pipeline.stream` is wired synchronously and emits the unprocessed input until the model is ready (~1–3 s on cold start), so users see live video the whole time. Model weights stream from `cdn.longpipe.dev` by default — no extra setup, and [self-hostable](https://longpipe.dev/docs/self-hosting).
 
-```ts
-await pipeline.setBackground({ blur: { sigma: 20 } })
-await pipeline.setBackground('https://example.com/other-bg.jpg')
-```
+## Why Longpipe
 
-## Performance presets
+- **One model, many effects.** Backgrounds, touch-up, and auto-reframe all run off a single shared encoder pass — enabling another effect doesn't cost another inference. Audio denoise runs in a parallel `AudioWorklet`, off the GPU entirely. ([Architecture](https://longpipe.dev/docs/architecture))
+- **More accurate *and* faster than the open-source alternatives.** Custom-trained models running as pure WebGPU/WebGL shaders in a zero-copy, fully-GPU pipeline — no general-purpose runtime, no CPU↔GPU round trips.
+- **Works everywhere.** Chromium, Firefox, and Safari (desktop + iOS). WebGPU when available, WebGL2 fallback, per-browser frame transport handled internally.
+- **Adapts to the device.** Five model presets (`xs`–`xl`); autotune benchmarks the actual device at init and an adaptive controller swaps presets at runtime to hold 30 fps — on 10-year-old netbooks and the latest MacBook Pro alike.
+- **Production pedigree.** Built by the founder of [Vectorly](https://www.crunchbase.com/organization/vectorly), a commercial effects SDK acquired in 2021 — a ground-up redesign for the WebGPU era.
 
-```ts
-new EffectsPipeline(stream, { preset: 'auto' })       // benchmarks at init (default)
-new EffectsPipeline(stream, { preset: 'fast' })       // small model
-new EffectsPipeline(stream, { preset: 'balanced' })   // medium model
-new EffectsPipeline(stream, { preset: 'quality' })    // xl model
-```
+### Best models on the web
 
-`'auto'` runs a microbenchmark at init and picks the largest preset that fits the per-frame budget on the current device, **topping out at `large`** — `xl` is the explicit opt-in flagship (`'quality'` or a manual config). While `'auto'` is in effect an adaptive controller polls FPS / model time and swaps preset up or down as conditions change. Explicit preset choices (`'fast'`, `'balanced'`, `'quality'`, or a manual config) are always respected and never auto-overridden.
+Longpipe's hydranet architecture, custom model-specific shaders, zero-memory-copy pipeline, developer friendly interface and 100% MIT open source license mean that 🐲 Longpipe pareto dominates on performance, model quality and ease of use.
 
-You can also pass a manual config:
+<img width="958" height="652" alt="lp-performance" src="https://github.com/user-attachments/assets/792b9f63-e8fd-4d30-aa90-9c92b6a0430c" />
 
-```ts
-new EffectsPipeline(stream, {
-  preset: { model: 'large', dtype: 'f32', resolution: { w: 640, h: 400 }, skipFrames: 0 },
-})
-```
 
-## Audio denoise
+## Documentation
 
-Real-time speech denoising runs as a **separate AudioWorklet pipeline** on the audio render thread — independent of the video/GPU pipeline, in parallel. Pass `audio: 'denoise'` and the input stream's mic is cleaned in place; `pipeline.stream` carries the denoised audio track.
+Full docs live at **[longpipe.dev/docs](https://longpipe.dev/docs)**:
 
-```ts
-const stream = await navigator.mediaDevices.getUserMedia({ video: true, audio: true })
-new EffectsPipeline(stream, { audio: 'denoise' })   // auto-picks a model for the device
-```
-
-Three models back three tiers. `auto` (the default) runs a tiny **weight-free probe** at init — it times the real network on this device *without* downloading the multi-MB weights — and picks the best fit:
-
-| `model`   | tier | what it is |
-|-----------|------|------------|
-| `dfn`     | high | DeepFilterNet3, full f32 — best quality |
-| `dfnint8` | mid  | DeepFilterNet3 with int8 GRUs — smaller download, faster on weak hardware |
-| `rnnoise` | low  | classic RNNoise — tiny, and the fallback where wasm SIMD is unavailable |
-
-```ts
-// force a specific model or tier instead of probing
-new EffectsPipeline(stream, { audio: { denoise: { model: 'dfn' } } })
-
-interface DenoiseOptions {
-  model?:          'auto' | 'high' | 'mid' | 'low' | 'rnnoise' | 'dfn' | 'dfnint8'  // default 'auto'
-  postFilterBeta?: number    // DFN post-filter, suppresses residual noise (default 0.03)
-  gruLeak?:        number     // DFN recurrent-drift bound (default 0.995)
-  enabled?:        boolean    // start denoising vs. passthrough (default true)
-}
-```
-
-Toggle, reconfigure, and inspect at runtime:
-
-```ts
-pipeline.setDenoise(false)                      // passthrough (cheap to re-enable)
-pipeline.setDenoise({ postFilterBeta: 0.05 })   // tweak DFN params live
-pipeline.getAudioStats()                        // { model, p50Ms, p95Ms, latencyMs, sampleRate, active }
-```
-
-The denoiser handles sample-rate conversion internally (it runs at 48 kHz and resamples when the device can't). `pipeline.ready` resolves on the video first frame and does **not** wait on audio — denoising joins asynchronously, with the mic passing through until it's live.
-
-## Other options
-
-```ts
-new EffectsPipeline(stream, {
-  background:       'blur',
-  touchup:          { strength: 0.6 },             // skin smoothing; omit to disable
-  reframe:          true,                          // auto-frame the subject; omit to disable
-  preset:           'auto',
-  adaptive:         true,                          // default; only applies when preset is 'auto'
-  audio:            'denoise',                 // 'passthrough' | 'drop' | 'denoise' | { denoise: {...} }
-  outputResolution: { w: 1280, h: 720 },           // default: matches the input video track
-  weightsBaseUrl:   'https://your-cdn/longpipe/',  // default: cdn.longpipe.dev
-  enabled:          true,                          // false = pass input through unchanged
-  onReady:          () => console.log('live'),
-  onError:          (err) => console.error(err),
-})
-```
-
-Toggle the effect on/off without tearing anything down (cheap to re-enable):
-
-```ts
-pipeline.setEnabled(false)
-```
-
-Tear it all down:
-
-```ts
-pipeline.destroy()
-```
-
-## Self-hosting weights
-
-By default Longpipe fetches model weights from `https://cdn.longpipe.dev/models/v/0.0.5/`. You can browse the available files, sizes, and SHA-256 hashes at [cdn.longpipe.dev/models/v/0.0.5/index.html](https://cdn.longpipe.dev/models/v/0.0.5/index.html) (machine-readable list at [manifest.json](https://cdn.longpipe.dev/models/v/0.0.5/manifest.json)).
-
-To serve them yourself, mirror the files under any base URL with the same `model_${name}.bin` naming convention and pass it via `weightsBaseUrl`:
-
-```ts
-new EffectsPipeline(stream, {
-  weightsBaseUrl: 'https://your-cdn.example.com/longpipe-weights/',
-})
-```
-
-If you use touch-up, three more files must sit alongside the tier weights — they're fetched from the same base URL the first time the effect is enabled: `model_landmark_mesh.bin` (and `.f16.bin`), `face_topology.json`, and `weight_mask.png`. Miss these and matting still works, but touch-up fails to enable. Auto-reframe needs nothing extra.
-
-## Browser support
-
-Works on Chromium (Chrome, Edge), Firefox, and Safari (desktop and iOS). WebGPU is used when available; WebGL2 is the fallback. Longpipe picks the optimal video frame transport for each browser internally — `MediaStreamTrackProcessor`, `transferControlToOffscreen` + `captureStream`, or an `ImageBitmap` shuttle as universal fallback — all invisible to the caller.
-
-
-
-## Face touch-up
-
-UV-space skin smoothing (the "beautify" filter). It runs **in parallel with any background effect** — both features share one encoder pass per frame, and the retouched frame feeds the background compositor.
-
-```ts
-new EffectsPipeline(stream, {
-  background: 'blur',
-  touchup: { strength: 0.6 },   // presence enables it
-})
-
-await pipeline.setTouchup({ strength: 0.8, style: 'bilateral' })   // live update
-await pipeline.setTouchup(null)                                    // disable
-```
-
-```ts
-interface TouchupOptions {
-  strength?: number   // 0..1 — blend of smoothed skin over original (default 0.6)
-  amount?:   number   // smoothing radius (default 8)
-  detail?:   number   // high-frequency (pore texture) keep — freq-sep only (default 0.35)
-  style?:    'freq-sep' | 'bilateral'   // default 'freq-sep'
-}
-```
-
-How it works: a 5-point face-keypoint head rides the matting encoder (no second encoder pass); its heatmaps are decoded on the GPU into a face box; a 478-point landmark model runs on the face crop; the face is unwrapped into a pose-independent UV atlas, smoothed there, and composited back through a per-region weight mask. Eyes, brows, and lips are protected and stay pixel-exact; `'freq-sep'` smooths while keeping pore-level texture, `'bilateral'` is a stronger edge-preserving filter. The whole chain is GPU-resident — no per-frame readback.
-
-On first enable the landmark weights (`model_landmark_mesh.bin`, ~2.6 MB) and two small static assets are fetched from `weightsBaseUrl`, then cached; parameter updates are instant. With no face in frame the effect passes through cleanly.
-
-Everyone in frame gets smoothed — up to **four faces** on the `medium`, `large` and `xl` presets. (`small` and `xs` decode keypoints on a coarser grid that can't reliably separate a second face, so they smooth one.) There's no per-face API and nothing to configure; the faces share a single atlas, so the smoothing passes run once regardless of how many people are in shot.
-
-## Auto-reframe
-
-Crop and zoom the output so the subject is framed — the "centre stage" behaviour from Meet, Zoom and Teams. It rides the same face keypoints as touch-up, so it needs **no extra model and no extra assets**: turning it on costs a crop, not an inference.
-
-```ts
-new EffectsPipeline(stream, { reframe: true })                 // auto-follow, all defaults
-new EffectsPipeline(stream, { reframe: { auto: false } })      // solve once, then freeze
-
-await pipeline.setReframe({ zoom: 1.6 })   // enable / update live
-await pipeline.reframe()                   // manual mode: re-solve now
-await pipeline.setReframe(null)            // disable
-```
-
-```ts
-type ReframeConfig = boolean | ReframeOptions
-
-interface ReframeOptions {
-  zoom?:    number   // crop = frame / zoom (default 1.35; relaxes toward 1 as needed)
-  gravity?: number   // 0..1 pull toward the subject (default 0.5; 1 would centre it exactly)
-  margin?:  number   // keep-out space around the face (default 0.04)
-  auto?:    boolean | { deadband?: number; ease?: number }   // default true
-}
-```
-
-Two modes, because apps ship both. **auto** (default) follows the subject with a deadband — it holds still, makes a deliberate move when they've actually gone somewhere, then holds again; continuous tracking reads as swimming, because a head moves constantly even when a person doesn't. **manual** (`auto: false`) frames once and freezes until you call `pipeline.reframe()` — the mode for a reframe button in your own UI.
-
-The crop must stay inside the frame *and* contain the subject, which has a useful consequence: if you sit hard in a corner, auto-reframe does nothing at all, because containing you would need a crop the size of the whole frame. That falls out of the constraints rather than being a special case, and it matches what the commercial implementations do.
-
-It reframes **you**, not your background: with a virtual background the image stays put while you zoom (it's a backdrop behind you, not part of the shot), while a *blurred* background reframes with you, since that is your real room. Neither needs configuring. Auto-reframe frames one subject — the largest face — and the choice is sticky, so two people side by side won't make the camera flip back and forth. That's the opposite of touch-up, which smooths everyone: framing is a choice about whose shot it is; smoothing isn't.
-
-## How it works
-
-Two layers:
-
-- **Model** (`src/model/`) — a shared EfficientNet-Lite encoder (lite0; lite3 on the `xl` tier) feeding multiple heads: a U-Net matting decoder (sharpened at higher resolution by a lightweight U-Net *wrapper*), a 5-point face-keypoint head (drives touch-up and auto-reframe), and a small optical-flow head that provides temporal stability — the renderer warps the previous outputs along the predicted flow and blends them through a flow-gated stabilizer, so the matte is steady without any recurrent state in the network. A separate 478-point landmark model runs on the detected face crop. Written as TypeScript op classes — each layer is a class; weights load as binary tensors at init; the backend (WebGPU or WebGL2) is injected at construction. BatchNorm is fused into conv weights at export — there is no BN op at inference.
-- **Pipeline** (`src/pipeline/`) — capability detection, per-browser frame transport selection, worker spawn, audio passthrough, autotune, and the adaptive controller. Designed to absorb the browser/codec/canvas plumbing complexity so consumers don't have to. 
-
-Five trained presets cover the hardware range. "Resolution" is the model's working (canvas) resolution; the encoder runs at a lower internal resolution and the U-Net wrapper refines back up to this.
-
-| Preset | Resolution | Encoder      | Decoder  | Skip frames |
-|--------|------------|--------------|----------|-------------|
-| xl     | 1280×768   | full (lite3) | 2× ch    | 0           |
-| large  | 640×400    | full (lite0) | standard | 0           |
-| medium | 512×320    | full (lite0) | standard | 1           |
-| small  | 384×224    | small        | standard | 1           |
-| xs     | 384×240    | small        | standard | 2           |
-
-`Skip frames` is how many input frames the model sits out between runs — the compositor still renders every frame using the most recent alpha matte. Autotune picks one of these at init based on a microbenchmark of the actual network on the actual device.
-
-## Development
-
-Training scripts, fixture generation, and the weight export pipeline are not yet documented here — coming soon.
+- [Getting started](https://longpipe.dev/docs/getting-started)
+- [API reference](https://longpipe.dev/docs/api) — all `EffectsPipeline` options and methods
+- [Architecture](https://longpipe.dev/docs/architecture) — the hydranet: one encoder, many heads
+- [Backgrounds](https://longpipe.dev/docs/backgrounds) · [Touch-up](https://longpipe.dev/docs/touchup) · [Auto-reframe](https://longpipe.dev/docs/reframe) · [Audio denoise](https://longpipe.dev/docs/audio)
+- [Presets & autotune](https://longpipe.dev/docs/presets)
+- [Model layer](https://longpipe.dev/docs/model-layer) — drive the model directly, without the pipeline
+- [Self-hosting weights](https://longpipe.dev/docs/self-hosting)
 
+Machine-readable for LLMs and agents: [`llms.txt`](https://longpipe.dev/llms.txt) · [`llms-full.txt`](https://longpipe.dev/llms-full.txt)
 
 ## Roadmap
 
@@ -292,25 +71,8 @@ Training scripts, fixture generation, and the weight export pipeline are not yet
 - [ ] AR effects
 - [ ] Lighting correction
 
-
 ## License
 
-SDK source code is MIT-licensed — see [LICENSE](LICENSE).
-
-The pre-trained model weights distributed via `cdn.longpipe.dev` (and any mirrored copies) are also MIT-licensed — see [WEIGHTS_LICENSE](WEIGHTS_LICENSE). You're free to self-host them inside other open-source or commercial projects under MIT-compatible terms.
-
-### Training data and pretrained weights
-
-Longpipe's pre-trained weights are released under MIT (see [WEIGHTS_LICENSE](WEIGHTS_LICENSE)). They were trained on:
-
-- [P3M-10k](https://github.com/JizhiziLi/P3M) — MIT
-- [AISegment Matting Human Datasets](https://www.kaggle.com/datasets/laurentmih/aisegmentcom-matting-human-datasets) — MIT ([upstream license](https://github.com/aisegmentcn/matting_human_datasets/blob/master/LICENSE))
-- [COCO](https://cocodataset.org) CC BY 4.0
-- [OpenImages](https://storage.googleapis.com/openimages/web/index.html) CC by 4.0
-- Synthetic images generated with [Z-Image](https://github.com/Tongyi-MAI/Z-Image) (text-to-image) — Apache 2.0
-- A custom dataset of short webcam videos collected via [Prolific](https://www.prolific.com/), from participants who gave explicit, informed consent for their footage to be used to train an open-source virtual-background model. Faces are blurred before training; the raw videos are stored privately (EU), never distributed, and deleted after the training window.
-- Pseudo-labels generated by [BiRefNet](https://github.com/ZhengPeng7/BiRefNet) (MIT), used as a teacher model
-
-The encoder is initialized from a pre-trained [EfficientNet-Lite](https://github.com/RangiLyu/EfficientNet-Lite) backbone — Apache 2.0.
+SDK source code **and** pre-trained model weights are MIT-licensed — see [LICENSE](LICENSE) and [WEIGHTS_LICENSE](WEIGHTS_LICENSE). You're free to self-host the weights inside other open-source or commercial projects under MIT-compatible terms.
 
-As is standard for trained models, the released weights are not a redistribution of any training image or video — they are published under MIT.
+The weights were trained on permissively-licensed public datasets (P3M-10k, AISegment, COCO, OpenImages), Z-Image synthetics, and a custom webcam dataset collected with participants' explicit, informed consent. The full dataset list and provenance details: [longpipe.dev/docs/licensing](https://longpipe.dev/docs/licensing).

package/dist/audio.cjs

@@ -0,0 +1,33 @@
+'use strict';var a={rnnoise:{wasm:"rnnoise.wasm",weights:null,scaleIn:32768,scaleOut:30517578125e-15,needsSimd:false,exports:{malloc:"malloc",create:"rn_create",process:"rn_process"}},dfn:{wasm:"dfn.wasm",weights:"dfn_weights.pack",scaleIn:1,scaleOut:1,needsSimd:true,exports:{malloc:"df_lite_malloc",create:"df_lite_create",process:"df_lite_process",setBeta:"df_lite_set_beta",setGruLeak:"df_lite_set_gru_leak"}},dfnint8:{wasm:"dfn.wasm",weights:"dfn_weights_int8.pack",scaleIn:1,scaleOut:1,needsSimd:true,exports:{malloc:"df_lite_malloc",create:"df_lite_create_i8",process:"df_lite_process",setBeta:"df_lite_set_beta",setGruLeak:"df_lite_set_gru_leak"}}},_={high:"dfn",mid:"dfnint8",low:"rnnoise"},b=["high","mid","low"];function x(t){return b.includes(t)}function p(t){return x(t)?_[t]:t}function m(t,e){return `${t.replace(/\/$/,"")}/${e}`}async function f(t){let e=await fetch(t);if(!e.ok)throw new Error(`audio asset fetch failed: ${e.status} ${t}`);return e.arrayBuffer()}async function g(t,e){let r=a[t],[s,o]=await Promise.all([f(m(e,r.wasm)),r.weights?f(m(e,r.weights)):Promise.resolve(null)]);return {wasmBytes:s,weights:o}}function l(){try{return WebAssembly.validate(new Uint8Array([0,97,115,109,1,0,0,0,1,5,1,96,0,1,123,3,2,1,0,10,22,1,20,0,253,12,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,11]))}catch{return  false}}var M=1.3,k=64,S=3,A=`
+self.onmessage = async (e) => {
+  const { wasmUrl, iters, rounds, budget } = e.data
+  try {
+    const res = await fetch(wasmUrl)
+    if (!res.ok) throw new Error(res.status + ' ' + wasmUrl)
+    const module = await WebAssembly.compile(await res.arrayBuffer())
+    const imports = {}
+    for (const im of WebAssembly.Module.imports(module)) {
+      ;(imports[im.module] = imports[im.module] || {})
+      if (im.kind === 'function') imports[im.module][im.name] = () => 0
+    }
+    const ex = new WebAssembly.Instance(module, imports).exports
+    const best = (fn) => {
+      fn(iters)                                   // warm up
+      let b = Infinity
+      for (let r = 0; r < rounds; r++) {
+        const t = performance.now()
+        fn(iters)
+        b = Math.min(b, (performance.now() - t) / iters)
+      }
+      return b
+    }
+    const dfnMs = best(ex.calibrate_f32)
+    const int8Ms = ex.calibrate_i8 ? best(ex.calibrate_i8) : Infinity
+    const model = dfnMs <= budget ? 'dfn' : int8Ms <= budget ? 'dfnint8' : 'rnnoise'
+    self.postMessage({ model, dfnMs, int8Ms })
+  } catch (err) {
+    self.postMessage({ error: (err && err.message) || String(err) })
+  }
+}
+`;async function w(t,e){if(!l())return "rnnoise";let r=`${t.replace(/\/$/,"")}/${a.dfn.wasm}`,s=null,o=null;try{s=URL.createObjectURL(new Blob([A],{type:"application/javascript"})),o=new Worker(s);let i=await new Promise((u,n)=>{o.onmessage=d=>u(d.data),o.onerror=d=>n(new Error(d.message||"probe worker error")),o.postMessage({wasmUrl:r,iters:k,rounds:S,budget:M});});if(i.error)throw new Error(i.error);return i.model??"rnnoise"}catch(i){return e?.(`audio tier probe failed (${i.message??String(i)}); using rnnoise`),"rnnoise"}finally{o?.terminate(),s&&URL.revokeObjectURL(s);}}var h='var b=Object.defineProperty;var y=(u,a,i)=>a in u?b(u,a,{enumerable:!0,configurable:!0,writable:!0,value:i}):u[a]=i;var r=(u,a,i)=>y(u,typeof a!="symbol"?a+"":a,i);var f={rnnoise:{wasm:"rnnoise.wasm",weights:null,scaleIn:32768,scaleOut:30517578125e-15,needsSimd:!1,exports:{malloc:"malloc",create:"rn_create",process:"rn_process"}},dfn:{wasm:"dfn.wasm",weights:"dfn_weights.pack",scaleIn:1,scaleOut:1,needsSimd:!0,exports:{malloc:"df_lite_malloc",create:"df_lite_create",process:"df_lite_process",setBeta:"df_lite_set_beta",setGruLeak:"df_lite_set_gru_leak"}},dfnint8:{wasm:"dfn.wasm",weights:"dfn_weights_int8.pack",scaleIn:1,scaleOut:1,needsSimd:!0,exports:{malloc:"df_lite_malloc",create:"df_lite_create_i8",process:"df_lite_process",setBeta:"df_lite_set_beta",setGruLeak:"df_lite_set_gru_leak"}}};var c=4096,m=class extends AudioWorkletProcessor{constructor(i){super();r(this,"cfg");r(this,"mem");r(this,"run");r(this,"state",0);r(this,"inPtr",0);r(this,"outPtr",0);r(this,"setBeta",null);r(this,"setGruLeak",null);r(this,"resample",!1);r(this,"rsPush",null);r(this,"inRs",0);r(this,"outRs",0);r(this,"rsDevInPtr",0);r(this,"rsOutPtr",0);r(this,"enabled");r(this,"model");r(this,"inAcc",new Float32Array(480));r(this,"inFill",0);r(this,"outRing",new Float32Array(4096));r(this,"outRead",0);r(this,"outWrite",0);r(this,"outCount",0);r(this,"primed",!1);r(this,"now");r(this,"times",new Float32Array(256));r(this,"tIdx",0);r(this,"hops",0);r(this,"_f32",null);let o=i.processorOptions;this.model=o.model,this.enabled=o.enabled,this.cfg=f[o.model];try{let t=this.instantiate(o.wasmBytes),e=n=>n?t[n]:void 0;this.mem=t.memory;let s=t[this.cfg.exports.malloc];if(this.run=t[this.cfg.exports.process],this.inPtr=s(480*4),this.outPtr=s(480*4),o.weights?this.state=t[this.cfg.exports.create](this.upload(s,o.weights),o.weights.byteLength):this.state=t[this.cfg.exports.create](),this.setBeta=e(this.cfg.exports.setBeta)??null,this.setGruLeak=e(this.cfg.exports.setGruLeak)??null,this.setBeta&&(o.postFilterBeta??0)>0&&this.setBeta(this.state,o.postFilterBeta),this.setGruLeak&&(o.gruLeak??1)<1&&this.setGruLeak(this.state,o.gruLeak),sampleRate!==48e3){let n=t.df_resampler_create;this.rsPush=t.df_resampler_push??null,n&&this.rsPush?(this.resample=!0,this.inRs=n(sampleRate,48e3),this.outRs=n(48e3,sampleRate),this.rsDevInPtr=s(256*4),this.rsOutPtr=s(c*4)):this.post({type:"error",message:`${o.model} can\'t resample ${sampleRate}\\u219248000 Hz; needs a 48 kHz AudioContext`})}}catch(t){let e=t;this.post({type:"error",message:`denoise init failed: ${e.message??String(t)}${e.stack?`\n`+e.stack:""}`})}this.now=typeof performance<"u"&&performance.now?()=>performance.now():null,this.port.onmessage=t=>{let e=t.data;e.type==="enabled"?this.enabled=e.value:e.type==="config"&&(this.setBeta&&e.postFilterBeta!=null&&this.setBeta(this.state,e.postFilterBeta),this.setGruLeak&&e.gruLeak!=null&&this.setGruLeak(this.state,e.gruLeak))},this.post({type:"ready"})}instantiate(i){var e;let o=new WebAssembly.Module(i),t={};for(let s of WebAssembly.Module.imports(o))t[e=s.module]??(t[e]={}),s.kind==="function"?t[s.module][s.name]=()=>0:s.kind==="memory"?t[s.module][s.name]=new WebAssembly.Memory({initial:256}):s.kind==="table"?t[s.module][s.name]=new WebAssembly.Table({initial:0,element:"anyfunc"}):s.kind==="global"&&(t[s.module][s.name]=new WebAssembly.Global({value:"i32",mutable:!1},0));return new WebAssembly.Instance(o,t).exports}upload(i,o){let t=new Uint8Array(o),e=i(t.length);return new Uint8Array(this.mem.buffer).set(t,e),e}heap(){return(!this._f32||this._f32.buffer!==this.mem.buffer)&&(this._f32=new Float32Array(this.mem.buffer)),this._f32}post(i){this.port.postMessage(i)}pushOut(i){this.outRing[this.outWrite]=i,this.outWrite=(this.outWrite+1)%this.outRing.length,this.outCount++}feed48k(i,o){for(let t=0;t<o;t++)this.inAcc[this.inFill++]=i[t],this.inFill===480&&(this.runHop(),this.inFill=0)}runHop(){let i=this.heap(),o=this.inPtr>>2;for(let e=0;e<480;e++)i[o+e]=this.inAcc[e]*this.cfg.scaleIn;let t=this.now?this.now():0;if(this.run(this.state,this.inPtr,this.outPtr),this.now&&(this.times[this.tIdx=this.tIdx+1&255]=this.now()-t),this.resample&&this.rsPush){let e=this.rsPush(this.outRs,this.outPtr,480,this.rsOutPtr,c);i=this.heap();let s=this.rsOutPtr>>2;for(let n=0;n<e;n++)this.pushOut(i[s+n])}else{let e=this.outPtr>>2;for(let s=0;s<480;s++)this.pushOut(i[e+s]*this.cfg.scaleOut)}++this.hops>=50&&(this.reportStats(),this.hops=0)}reportStats(){let i=Array.from(this.times).filter(e=>e>0).sort((e,s)=>e-s),o=e=>i.length?i[Math.min(i.length-1,Math.floor(e*i.length))]:null,t={model:this.model,p50Ms:this.now?o(.5):null,p95Ms:this.now?o(.95):null,latencyMs:this.outCount/sampleRate*1e3,active:this.enabled&&this.state!==0,sampleRate};this.post({type:"stats",stats:t})}process(i,o){let t=i[0]?.[0],e=o[0]?.[0];if(!e)return!0;let s=e.length;if(!t||!this.enabled||this.state===0)return t&&!this.enabled?e.set(t):e.fill(0),!0;if(this.resample&&this.rsPush){let n=this.heap(),g=this.rsDevInPtr>>2;for(let h=0;h<s;h++)n[g+h]=t[h];let d=this.rsPush(this.inRs,this.rsDevInPtr,s,this.rsOutPtr,c);n=this.heap(),this.feed48k(n.subarray(this.rsOutPtr>>2,(this.rsOutPtr>>2)+d),d)}else this.feed48k(t,s);if(!this.primed)if(this.outCount>=480)this.primed=!0;else return e.fill(0),!0;if(this.outCount>=s)for(let n=0;n<s;n++)e[n]=this.outRing[this.outRead],this.outRead=(this.outRead+1)%this.outRing.length,this.outCount--;else e.fill(0);return!0}};registerProcessor("denoise",m);\n';var y="https://cdn.longpipe.dev/models/v/0.0.5/";var E=.03,v=.995,c=class{constructor(e,r={}){this.opts=r;this.ctx=new AudioContext({sampleRate:48e3}),this.source=this.ctx.createMediaStreamSource(new MediaStream([e])),this.dest=this.ctx.createMediaStreamDestination(),this.outputTrack=this.dest.stream.getAudioTracks()[0],this.resume().catch(()=>{}),this.source.connect(this.dest),this.ready=this.init().then(()=>this.node!==null,s=>(this.opts.onError?.(`audio denoise init failed: ${s.message??String(s)}`),false));}opts;outputTrack;ready;ctx;source;dest;node=null;model=null;latestStats=null;destroyed=false;resolveExplicit(){let e=p(this.opts.model);return a[e].needsSimd&&!l()?(this.opts.onError?.(`${e} needs wasm SIMD; falling back to rnnoise`),"rnnoise"):e}async init(){let e=this.opts.weightsBaseUrl??y,r=(this.opts.model??"auto")==="auto"?await w(e,this.opts.onError):this.resolveExplicit();this.model=r;let s=await g(r,e);if(await this.ctx.audioWorklet.addModule(D()),this.destroyed)return;let o={model:r,wasmBytes:s.wasmBytes,weights:s.weights,enabled:this.opts.enabled??true,postFilterBeta:this.opts.postFilterBeta??E,gruLeak:this.opts.gruLeak??v},i=new AudioWorkletNode(this.ctx,"denoise",{numberOfInputs:1,numberOfOutputs:1,outputChannelCount:[1],processorOptions:o});i.port.onmessage=u=>{let n=u.data;n.type==="stats"?this.latestStats=n.stats:n.type==="error"?this.opts.onError?.(n.message):n.type==="ready"&&this.opts.onReady?.();},this.node=i,this.source.disconnect(this.dest),this.source.connect(i).connect(this.dest);}setEnabled(e){this.node?.port.postMessage({type:"enabled",value:e});}setConfig(e){this.node?.port.postMessage({type:"config",...e});}getStats(){return this.latestStats}async resume(){this.destroyed||this.ctx.state!=="suspended"||await this.ctx.resume();}destroy(){this.destroyed=true;try{this.node?.disconnect();}catch{}try{this.source.disconnect();}catch{}this.ctx.close();}};function D(){return URL.createObjectURL(new Blob([h],{type:"application/javascript"}))}
+exports.AudioDenoiser=c;

package/dist/audio.d.cts

@@ -0,0 +1,37 @@
+import { D as DenoiseModelOption, A as AudioStats } from './messages-DjW-szBq.cjs';
+export { a as DenoiseModel, b as DenoiseTier } from './messages-DjW-szBq.cjs';
+
+interface AudioDenoiserOptions {
+    model?: DenoiseModelOption;
+    weightsBaseUrl?: string;
+    postFilterBeta?: number;
+    gruLeak?: number;
+    enabled?: boolean;
+    onReady?: () => void;
+    onError?: (message: string) => void;
+}
+declare class AudioDenoiser {
+    private opts;
+    readonly outputTrack: MediaStreamTrack;
+    readonly ready: Promise<boolean>;
+    private ctx;
+    private source;
+    private dest;
+    private node;
+    private model;
+    private latestStats;
+    private destroyed;
+    constructor(inputTrack: MediaStreamTrack, opts?: AudioDenoiserOptions);
+    private resolveExplicit;
+    private init;
+    setEnabled(on: boolean): void;
+    setConfig(cfg: {
+        postFilterBeta?: number;
+        gruLeak?: number;
+    }): void;
+    getStats(): AudioStats | null;
+    resume(): Promise<void>;
+    destroy(): void;
+}
+
+export { AudioDenoiser, type AudioDenoiserOptions, AudioStats, DenoiseModelOption };

package/dist/audio.d.ts

@@ -0,0 +1,37 @@
+import { D as DenoiseModelOption, A as AudioStats } from './messages-DjW-szBq.js';
+export { a as DenoiseModel, b as DenoiseTier } from './messages-DjW-szBq.js';
+
+interface AudioDenoiserOptions {
+    model?: DenoiseModelOption;
+    weightsBaseUrl?: string;
+    postFilterBeta?: number;
+    gruLeak?: number;
+    enabled?: boolean;
+    onReady?: () => void;
+    onError?: (message: string) => void;
+}
+declare class AudioDenoiser {
+    private opts;
+    readonly outputTrack: MediaStreamTrack;
+    readonly ready: Promise<boolean>;
+    private ctx;
+    private source;
+    private dest;
+    private node;
+    private model;
+    private latestStats;
+    private destroyed;
+    constructor(inputTrack: MediaStreamTrack, opts?: AudioDenoiserOptions);
+    private resolveExplicit;
+    private init;
+    setEnabled(on: boolean): void;
+    setConfig(cfg: {
+        postFilterBeta?: number;
+        gruLeak?: number;
+    }): void;
+    getStats(): AudioStats | null;
+    resume(): Promise<void>;
+    destroy(): void;
+}
+
+export { AudioDenoiser, type AudioDenoiserOptions, AudioStats, DenoiseModelOption };

package/dist/audio.js

@@ -0,0 +1 @@
+export{b as AudioDenoiser}from'./chunk-O4OF2WRO.js';

package/dist/chunk-JM6V2XLC.js

@@ -0,0 +1,33 @@
+var a={rnnoise:{wasm:"rnnoise.wasm",weights:null,scaleIn:32768,scaleOut:30517578125e-15,needsSimd:false,exports:{malloc:"malloc",create:"rn_create",process:"rn_process"}},dfn:{wasm:"dfn.wasm",weights:"dfn_weights.pack",scaleIn:1,scaleOut:1,needsSimd:true,exports:{malloc:"df_lite_malloc",create:"df_lite_create",process:"df_lite_process",setBeta:"df_lite_set_beta",setGruLeak:"df_lite_set_gru_leak"}},dfnint8:{wasm:"dfn.wasm",weights:"dfn_weights_int8.pack",scaleIn:1,scaleOut:1,needsSimd:true,exports:{malloc:"df_lite_malloc",create:"df_lite_create_i8",process:"df_lite_process",setBeta:"df_lite_set_beta",setGruLeak:"df_lite_set_gru_leak"}}},_={high:"dfn",mid:"dfnint8",low:"rnnoise"},b=["high","mid","low"];function M(t){return b.includes(t)}function c(t){return M(t)?_[t]:t}function m(t,e){return `${t.replace(/\/$/,"")}/${e}`}async function p(t){let e=await fetch(t);if(!e.ok)throw new Error(`audio asset fetch failed: ${e.status} ${t}`);return e.arrayBuffer()}async function f(t,e){let r=a[t],[s,o]=await Promise.all([p(m(e,r.wasm)),r.weights?p(m(e,r.weights)):Promise.resolve(null)]);return {wasmBytes:s,weights:o}}function l(){try{return WebAssembly.validate(new Uint8Array([0,97,115,109,1,0,0,0,1,5,1,96,0,1,123,3,2,1,0,10,22,1,20,0,253,12,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,11]))}catch{return  false}}var k=1.3,x=64,S=3,R=`
+self.onmessage = async (e) => {
+  const { wasmUrl, iters, rounds, budget } = e.data
+  try {
+    const res = await fetch(wasmUrl)
+    if (!res.ok) throw new Error(res.status + ' ' + wasmUrl)
+    const module = await WebAssembly.compile(await res.arrayBuffer())
+    const imports = {}
+    for (const im of WebAssembly.Module.imports(module)) {
+      ;(imports[im.module] = imports[im.module] || {})
+      if (im.kind === 'function') imports[im.module][im.name] = () => 0
+    }
+    const ex = new WebAssembly.Instance(module, imports).exports
+    const best = (fn) => {
+      fn(iters)                                   // warm up
+      let b = Infinity
+      for (let r = 0; r < rounds; r++) {
+        const t = performance.now()
+        fn(iters)
+        b = Math.min(b, (performance.now() - t) / iters)
+      }
+      return b
+    }
+    const dfnMs = best(ex.calibrate_f32)
+    const int8Ms = ex.calibrate_i8 ? best(ex.calibrate_i8) : Infinity
+    const model = dfnMs <= budget ? 'dfn' : int8Ms <= budget ? 'dfnint8' : 'rnnoise'
+    self.postMessage({ model, dfnMs, int8Ms })
+  } catch (err) {
+    self.postMessage({ error: (err && err.message) || String(err) })
+  }
+}
+`;async function g(t,e){if(!l())return "rnnoise";let r=`${t.replace(/\/$/,"")}/${a.dfn.wasm}`,s=null,o=null;try{s=URL.createObjectURL(new Blob([R],{type:"application/javascript"})),o=new Worker(s);let i=await new Promise((h,n)=>{o.onmessage=u=>h(u.data),o.onerror=u=>n(new Error(u.message||"probe worker error")),o.postMessage({wasmUrl:r,iters:x,rounds:S,budget:k});});if(i.error)throw new Error(i.error);return i.model??"rnnoise"}catch(i){return e?.(`audio tier probe failed (${i.message??String(i)}); using rnnoise`),"rnnoise"}finally{o?.terminate(),s&&URL.revokeObjectURL(s);}}var d='var b=Object.defineProperty;var y=(u,a,i)=>a in u?b(u,a,{enumerable:!0,configurable:!0,writable:!0,value:i}):u[a]=i;var r=(u,a,i)=>y(u,typeof a!="symbol"?a+"":a,i);var f={rnnoise:{wasm:"rnnoise.wasm",weights:null,scaleIn:32768,scaleOut:30517578125e-15,needsSimd:!1,exports:{malloc:"malloc",create:"rn_create",process:"rn_process"}},dfn:{wasm:"dfn.wasm",weights:"dfn_weights.pack",scaleIn:1,scaleOut:1,needsSimd:!0,exports:{malloc:"df_lite_malloc",create:"df_lite_create",process:"df_lite_process",setBeta:"df_lite_set_beta",setGruLeak:"df_lite_set_gru_leak"}},dfnint8:{wasm:"dfn.wasm",weights:"dfn_weights_int8.pack",scaleIn:1,scaleOut:1,needsSimd:!0,exports:{malloc:"df_lite_malloc",create:"df_lite_create_i8",process:"df_lite_process",setBeta:"df_lite_set_beta",setGruLeak:"df_lite_set_gru_leak"}}};var c=4096,m=class extends AudioWorkletProcessor{constructor(i){super();r(this,"cfg");r(this,"mem");r(this,"run");r(this,"state",0);r(this,"inPtr",0);r(this,"outPtr",0);r(this,"setBeta",null);r(this,"setGruLeak",null);r(this,"resample",!1);r(this,"rsPush",null);r(this,"inRs",0);r(this,"outRs",0);r(this,"rsDevInPtr",0);r(this,"rsOutPtr",0);r(this,"enabled");r(this,"model");r(this,"inAcc",new Float32Array(480));r(this,"inFill",0);r(this,"outRing",new Float32Array(4096));r(this,"outRead",0);r(this,"outWrite",0);r(this,"outCount",0);r(this,"primed",!1);r(this,"now");r(this,"times",new Float32Array(256));r(this,"tIdx",0);r(this,"hops",0);r(this,"_f32",null);let o=i.processorOptions;this.model=o.model,this.enabled=o.enabled,this.cfg=f[o.model];try{let t=this.instantiate(o.wasmBytes),e=n=>n?t[n]:void 0;this.mem=t.memory;let s=t[this.cfg.exports.malloc];if(this.run=t[this.cfg.exports.process],this.inPtr=s(480*4),this.outPtr=s(480*4),o.weights?this.state=t[this.cfg.exports.create](this.upload(s,o.weights),o.weights.byteLength):this.state=t[this.cfg.exports.create](),this.setBeta=e(this.cfg.exports.setBeta)??null,this.setGruLeak=e(this.cfg.exports.setGruLeak)??null,this.setBeta&&(o.postFilterBeta??0)>0&&this.setBeta(this.state,o.postFilterBeta),this.setGruLeak&&(o.gruLeak??1)<1&&this.setGruLeak(this.state,o.gruLeak),sampleRate!==48e3){let n=t.df_resampler_create;this.rsPush=t.df_resampler_push??null,n&&this.rsPush?(this.resample=!0,this.inRs=n(sampleRate,48e3),this.outRs=n(48e3,sampleRate),this.rsDevInPtr=s(256*4),this.rsOutPtr=s(c*4)):this.post({type:"error",message:`${o.model} can\'t resample ${sampleRate}\\u219248000 Hz; needs a 48 kHz AudioContext`})}}catch(t){let e=t;this.post({type:"error",message:`denoise init failed: ${e.message??String(t)}${e.stack?`\n`+e.stack:""}`})}this.now=typeof performance<"u"&&performance.now?()=>performance.now():null,this.port.onmessage=t=>{let e=t.data;e.type==="enabled"?this.enabled=e.value:e.type==="config"&&(this.setBeta&&e.postFilterBeta!=null&&this.setBeta(this.state,e.postFilterBeta),this.setGruLeak&&e.gruLeak!=null&&this.setGruLeak(this.state,e.gruLeak))},this.post({type:"ready"})}instantiate(i){var e;let o=new WebAssembly.Module(i),t={};for(let s of WebAssembly.Module.imports(o))t[e=s.module]??(t[e]={}),s.kind==="function"?t[s.module][s.name]=()=>0:s.kind==="memory"?t[s.module][s.name]=new WebAssembly.Memory({initial:256}):s.kind==="table"?t[s.module][s.name]=new WebAssembly.Table({initial:0,element:"anyfunc"}):s.kind==="global"&&(t[s.module][s.name]=new WebAssembly.Global({value:"i32",mutable:!1},0));return new WebAssembly.Instance(o,t).exports}upload(i,o){let t=new Uint8Array(o),e=i(t.length);return new Uint8Array(this.mem.buffer).set(t,e),e}heap(){return(!this._f32||this._f32.buffer!==this.mem.buffer)&&(this._f32=new Float32Array(this.mem.buffer)),this._f32}post(i){this.port.postMessage(i)}pushOut(i){this.outRing[this.outWrite]=i,this.outWrite=(this.outWrite+1)%this.outRing.length,this.outCount++}feed48k(i,o){for(let t=0;t<o;t++)this.inAcc[this.inFill++]=i[t],this.inFill===480&&(this.runHop(),this.inFill=0)}runHop(){let i=this.heap(),o=this.inPtr>>2;for(let e=0;e<480;e++)i[o+e]=this.inAcc[e]*this.cfg.scaleIn;let t=this.now?this.now():0;if(this.run(this.state,this.inPtr,this.outPtr),this.now&&(this.times[this.tIdx=this.tIdx+1&255]=this.now()-t),this.resample&&this.rsPush){let e=this.rsPush(this.outRs,this.outPtr,480,this.rsOutPtr,c);i=this.heap();let s=this.rsOutPtr>>2;for(let n=0;n<e;n++)this.pushOut(i[s+n])}else{let e=this.outPtr>>2;for(let s=0;s<480;s++)this.pushOut(i[e+s]*this.cfg.scaleOut)}++this.hops>=50&&(this.reportStats(),this.hops=0)}reportStats(){let i=Array.from(this.times).filter(e=>e>0).sort((e,s)=>e-s),o=e=>i.length?i[Math.min(i.length-1,Math.floor(e*i.length))]:null,t={model:this.model,p50Ms:this.now?o(.5):null,p95Ms:this.now?o(.95):null,latencyMs:this.outCount/sampleRate*1e3,active:this.enabled&&this.state!==0,sampleRate};this.post({type:"stats",stats:t})}process(i,o){let t=i[0]?.[0],e=o[0]?.[0];if(!e)return!0;let s=e.length;if(!t||!this.enabled||this.state===0)return t&&!this.enabled?e.set(t):e.fill(0),!0;if(this.resample&&this.rsPush){let n=this.heap(),g=this.rsDevInPtr>>2;for(let h=0;h<s;h++)n[g+h]=t[h];let d=this.rsPush(this.inRs,this.rsDevInPtr,s,this.rsOutPtr,c);n=this.heap(),this.feed48k(n.subarray(this.rsOutPtr>>2,(this.rsOutPtr>>2)+d),d)}else this.feed48k(t,s);if(!this.primed)if(this.outCount>=480)this.primed=!0;else return e.fill(0),!0;if(this.outCount>=s)for(let n=0;n<s;n++)e[n]=this.outRing[this.outRead],this.outRead=(this.outRead+1)%this.outRing.length,this.outCount--;else e.fill(0);return!0}};registerProcessor("denoise",m);\n';var w="https://cdn.longpipe.dev/models/v/0.0.5/";var E=.03,v=.995,y=class{constructor(e,r={}){this.opts=r;this.ctx=new AudioContext({sampleRate:48e3}),this.source=this.ctx.createMediaStreamSource(new MediaStream([e])),this.dest=this.ctx.createMediaStreamDestination(),this.outputTrack=this.dest.stream.getAudioTracks()[0],this.source.connect(this.dest),this.ready=this.init().then(()=>this.node!==null,s=>(this.opts.onError?.(`audio denoise init failed: ${s.message??String(s)}`),false));}opts;outputTrack;ready;ctx;source;dest;node=null;model=null;latestStats=null;destroyed=false;resolveExplicit(){let e=c(this.opts.model);return a[e].needsSimd&&!l()?(this.opts.onError?.(`${e} needs wasm SIMD; falling back to rnnoise`),"rnnoise"):e}async init(){let e=this.opts.weightsBaseUrl??w,r=(this.opts.model??"auto")==="auto"?await g(e,this.opts.onError):this.resolveExplicit();this.model=r;let s=await f(r,e);if(await this.ctx.audioWorklet.addModule(L()),this.destroyed)return;let o={model:r,wasmBytes:s.wasmBytes,weights:s.weights,enabled:this.opts.enabled??true,postFilterBeta:this.opts.postFilterBeta??E,gruLeak:this.opts.gruLeak??v},i=new AudioWorkletNode(this.ctx,"denoise",{numberOfInputs:1,numberOfOutputs:1,outputChannelCount:[1],processorOptions:o});i.port.onmessage=h=>{let n=h.data;n.type==="stats"?this.latestStats=n.stats:n.type==="error"?this.opts.onError?.(n.message):n.type==="ready"&&this.opts.onReady?.();},this.node=i,this.source.disconnect(this.dest),this.source.connect(i).connect(this.dest);}setEnabled(e){this.node?.port.postMessage({type:"enabled",value:e});}setConfig(e){this.node?.port.postMessage({type:"config",...e});}getStats(){return this.latestStats}destroy(){this.destroyed=true;try{this.node?.disconnect();}catch{}try{this.source.disconnect();}catch{}this.ctx.close();}};function L(){return URL.createObjectURL(new Blob([d],{type:"application/javascript"}))}
+export{w as a,y as b};