CoolFace
Apppublic

WJ88/NVIDIA-Parakeet-TDT-0.6B-v2-INT8-Real-Time-Mic-Transcription

sourceHugging Faceapache-2.0updated 1y agoView on Hugging Face
16likes
App README

Real-time English speech-to-text in your browser — no GPU required.

This Space runs the 600 M-parameter `nvidia/parakeet-tdt-0.6b-v2` model that fits comfortably on the CPU Basic (2 vCPU) tier.

  1. 1.Click “Record”
  2. 2.Allow microphone access and start speaking.
  3. 3.Watch live text appear in the Transcription box.

Stalled UI? Refresh the browser tab — this fully restarts the Space and clears any stuck threads.

TechniqueWhy it matters
`OMP_NUM_THREADS=2` & `torch.set_num_threads(2)`Matches the 2 vCPUs for optimal throughput
FBGEMM backendFastest kernels on x86
4-second streaming windowLow latency & small memory footprint
Gradio `stream_every=0.5`Updates the transcript twice per second for real-time feel
ItemLicence
Demo code (this repo)Apache-2.0
Model weights – nvidia/parakeet-tdt-0.6b-v2CC-BY-4.0 (© NVIDIA)

If you redistribute transcripts or fine-tuned weights, please retain the CC-BY-4.0 attribution notice.