| --- |
| license: mit |
| tags: |
| - webgpu |
| - wasm |
| - in-browser |
| - asr |
| - tts |
| - vad |
| - speech |
| - fluidaudio |
| --- |
| |
| # FluidAudio Web — raw model weights |
|
|
| Model weights for **[fluidaudio-web](https://github.com/FluidInference/fluidaudio-web)**: |
| fully in-browser inference (WebGPU + WebAssembly) for FluidAudio's speech models. |
|
|
| These are **raw weights only — no ONNX, no onnxruntime.** Each engine is a |
| hand-written forward pass (WebGPU/WASM/JS) that loads these tensors directly and, |
| where applicable, dequantizes them in-shader. Weights are extracted from the |
| source models and parity-verified against the originals before publishing. |
|
|
| ## Layout (one folder per engine) |
|
|
| | folder | model | weights | |
| |---|---|---| |
| | `vad/` | Silero VAD v5 (16 kHz) | fp32 | |
| | `parakeet/` | Parakeet TDT 0.6B v3 | int8 encoder + fp32 decoder/joint | |
| | `nemotron/` | Nemotron 3.5 streaming ASR 0.6B (40 langs) | int8 encoder + fp32 decoder | |
| | `eou/` | Parakeet EOU 120M (end-of-utterance ASR) | fp16 encoder + fp32 decoder | |
| | `voicechat-stt/` | NVIDIA VoiceChat-11B user-transcription chain (609M causal FastConformer + RNNT) | fp16 encoder + fp32 decoder | |
| | `whisper/` | Whisper base | fp32 encoder + fp32 decoder | |
| | `sortformer/` | Sortformer streaming diarization 4spk v2.1 | int8 encoder + fp32 head | |
| | `kokoro/` | Kokoro TTS 82M — English | fp32 | |
| | `kokoro-zh/` | Kokoro TTS 82M — Chinese | fp32 | |
|
|
| ## Format |
|
|
| - `*.bin` — concatenated little-endian tensors. |
| - `manifest.json` — `name -> { dims, offset, len }` (offset/len in **elements**, not bytes). |
|
|
| Load: fetch the `.bin`, slice each tensor per the manifest. Extraction scripts and |
| the forward passes live in the fluidaudio-web repo (`scripts/`, `src/engines/`). |
|
|
| Quantized engines additionally carry per-tensor scales / palettes (documented per |
| folder) for in-shader dequant. |
|
|
