README.md

metadata

title: LambdaAudio v6
colorFrom: blue
colorTo: indigo
sdk: gradio
sdk_version: 6.27.0
python_version: '3.12'
app_file: app.py
pinned: false
models:
  - KeisukeMiyamoto/lambda-audio-v6
short_description: Japanese speech recognition with Silero VAD

LambdaAudio v6

A simple Japanese speech recognition demo powered by LambdaAudio v6. Upload or record Japanese audio, then select 文字起こし (maximum file size: 200 MB). Audio up to 30 seconds is transcribed directly. Longer recordings are segmented with Silero VAD into speech clips of at most 30 seconds, using a 3-second silence threshold, and transcribed in batches of four. Results appear in chronological order as each batch finishes.

This Space uses ZeroGPU, with one recording at a time and a separate GPU allocation for each batch. GPU availability and usage limits depend on the visitor's Hugging Face account.

The FP16 Safetensors checkpoint is loaded into FP32 model parameters for inference. Audio and transcripts are not added to a dataset. Gradio temporary files are cleaned hourly once they are at least one hour old.

Deployment

Select ZeroGPU hardware and set HF_TOKEN as a Space secret. Upload this file as README.md, src/space/app.py as app.py, src/space/requirements.txt as requirements.txt, and the required Python modules under src/. Never upload .env or local training artifacts.