S3-S2ST: Semantically Supervised Streaming Speech-to-Speech Translation
Paper • 2610.36777 • Published • 1
S3-S2ST
4B models for Semantically Supervised Streaming Speech-to-Speech Translation.
This repository provides two variants, each with its matching model weights and complete rendering assets:
| Directory | Model |
|---|---|
streaming/ |
Streaming S2ST with READ/WRITE gate |
offline/ |
Offline S2ST |
Download and install
From the code checkout after installing the inference dependencies:
python tools/download_model.py --mode streaming --repo-id zhaoyx39/S3-S2ST --subfolder streaming
# Or the offline variant:
python tools/download_model.py --mode offline --repo-id zhaoyx39/S3-S2ST --subfolder offline
The installer downloads only the requested variant, verifies SHA256 and installs weights under models/. Follow the code README for audio preparation and inference commands. The model tensor values match the paper exports; paths are portable and no optimizer state, training data or experiment logs are included. Third-party terms are retained in each variant's licenses/ directory and THIRD_PARTY_NOTICES.md.