Lisper ZeroGPU

Server-side Gemma 4 audio analysis for users whose browser cannot comfortably run the WebGPU model.

The currently validated fine-tuned Lisper model is Gemma 4 E2B. E4B and 31B are future model targets and should be deployed as separate revisions after training/eval.

Browser recorder

No browser recording ready. This path bypasses Gradio's microphone recorder.

Upload fallback

Speech clip upload

No uploaded clip ready. Use the browser recorder above, or upload an audio file here.

Expected text

Gemma response

Parsed JSON

Configured model: thomasjvu/lisper-gemma4-e2b-audio-full

Configured adapter: none

Adapter 4-bit load: False

Acoustic hint: True

Audio token alignment: True

ZeroGPU size: large

If this Space errors on private or gated models, add HF_TOKEN as a Space secret. For local development without downloading the model, set LISPER_ZERO_GPU_EAGER_LOAD=0.