fusorlabs/voice-call-wasm-web

A voice call's three models, compiled for the browser

Three fusor-ai bundles, the weights the browser runtimes of voice-call-wasm load to answer a voice call on the device's CPU: not one speech-to-speech model but three working together. Try it in the Space fusorlabs/voice-call-wasm.

Folder	Model	Source	Profile	License
parakeet/	Speech to text	moondream/parakeet-redux at 2bf128600aac4b16946f7ed8372e56117fe5e23b	ternary: the encoder's 1.58-bit weights as published (199 MB)	CC-BY-4.0, from NVIDIA's parakeet-tdt-0.6b-v3
lfm3-1.7b/	The reply	LiquidAI/LFM3-1.7B at 40ebe0aa6662c9d40389c2f5d565b6a2c6aebacd	q7: every large matrix 7-bit (1964 MB)	LFM Open License v1.0 (lfm3-1.7b/LICENSE)
phonon/	Text to speech	gradium/phonon-31m-en at 52433c46975bfa13144c56204c9f4ee1c5b148aa, voices from gradium/phonon-voices at 0ee17205f254b5abc43b266c8a0f215565c352f7	q7 (46 MB)	Apache-2.0

• Format: each folder's manifest.json lists every tensor and part; each part is named by its BLAKE3 hash, which the runtime checks as it loads.
• Reproduce it: just model in voice-call-wasm downloads the checkpoints and compiles the same bundles.

These are not PyTorch or safetensors checkpoints: they only load in the voice-call-wasm runtimes built with the same fusor-ai.