@katana-video/carrot-backends

Model-backed backends for Carrot 🥕, the browser-native media engine. Each backend is its own subpath import and plugs into the engine's kernel/deriver seams; heavy runtimes are optional peer dependencies, dynamic-imported on first use.

npm i @katana-video/carrot @katana-video/carrot-backends mediabunny
npm i @huggingface/transformers        # for /whisper and /translate
npm i @katana-video/web-bg-removal     # for /removebg

// engine-worker.ts
import { createWorkerHost } from '@katana-video/carrot';
import { registerDenoise } from '@katana-video/carrot-backends/denoise';
import { removebg } from '@katana-video/carrot-backends/removebg';
import { transcribe } from '@katana-video/carrot-backends/whisper';
import { translate } from '@katana-video/carrot-backends/translate';

createWorkerHost({
  init: () => registerDenoise(),
  derivers: { transcribe, translate, removebg },
});

| Import | Provides | Runtime | Weights | | --- | --- | --- | --- | | /denoise | registerDenoise() — the denoise effect: DeepFilterNet (f32/int8) or RNNoise, wasm, 48 kHz | none | cdn.katana.video (see WEIGHTS_LICENSE) | | /enhance | registerEnhance() — the enhance effect: VoiceFixer voice restoration on onnxruntime-web/WebGPU, 3 s chunks stitched with seam-aware equal-power crossfades; also enhanceBuffer, prepareEnhance, enhanceSupported | onnxruntime-web 1.23.2 (pinned CDN import) | cdn.katana.video, CC BY 4.0 (see WEIGHTS_LICENSE) | | /whisper | transcribe deriver — whisper tiny/base/small with word timestamps; also extractPcm16k, toVTT/toSRT | @huggingface/transformers | Hugging Face hub | | /translate | translate deriver — OPUS-MT, English pivot for exotic pairs; also TRANSLATE_LANGUAGES, resolveTranslationRoute | @huggingface/transformers | Hugging Face hub | | /removebg | removebg deriver — BiRefNet-lite (quality) / people-only (fast) mattes | @katana-video/web-bg-removal | web-bg-removal's CDN | | /weights | configureWeights({ baseUrl }) to mirror everything from one origin | — | — |

Full contract, weights layout and mirroring: docs/backends.md.

Tests

node test/fetch-denoise-vendor.mjs   # once: the denoise wasm + packs (~22 MB) into test/vendor/
pnpm test                            # denoise (Node), whisper + translate (download models on first run), chain

License

MIT. Redistributed model weights carry their own licenses — see WEIGHTS_LICENSE.