VoxCPM2 WebGPU

Speech synthesis entirely in the browser — WASM + WebGPU, no inference server.

Weight precision Weight quantization Weights source Base URL for config.json, tokenizer.json, model.safetensors

AudioVAE source

Upstream publishes the AudioVAE only as audiovae.pth, a Python pickle this build cannot read, and no .safetensors version is hosted anywhere public. Convert it once with the convert_audiovae example, then either point at a CORS-friendly host (a Hugging Face repo works — GitHub release assets do not) or just pick the local file: it never leaves this tab.

URL

Using the URL above.

Allow a software WebGPU adapter (SwiftShader / llvmpipe). Off by default: this is a 2.3 B-parameter model, so a CPU rasterizer takes hours and any timing it produces is meaningless. Only for correctness testing where no GPU is reachable.

Text

Options

CFG scale Diffusion steps Max patches

Stream audio while generating. Plays each chunk as it is produced rather than waiting for the whole utterance. Lowers time-to-first-audio; raises total work, because every chunk re-decodes the accumulated latent. Deterministic (VOXCPM_Z_ZERO): zero the diffusion sampler's noise so a run can be diffed numerically against native. Degrades the audio — measured on native, it drops voiced frames from ~57% to ~24% and stops the stop-head firing, so the model runs to max_len. Use it to compare numbers, not to judge quality.

Reference voice (voice cloning)

No reference — the model improvises a voice. Audio stays in this tab; nothing is uploaded.

Status

idle

Metrics

WebGPU adapter	—
Weight precision	—
GPU weights (measured)	—
Model load time	—
Generation time	—
Time to first audio	—
Chunks / underruns	—
Audio duration	—
RTF (gen / audio)	—
Peak amplitude	—
WASM heap	—

RTF < 1 means faster than realtime.

VoxCPM2 WebGPU

WebGPU is not available

Status

Metrics

Diagnostics