xdp-tts-service 0.1.1

Standalone TTS Flask service package for XDP

xdp-tts-service wheel packaging

This directory builds a standalone wheel for the TTS Flask service without changing the monorepo layout.

What it does

• Copies runtime files from text_to_speech/ into a temporary staging directory during build.
• Bundles a sanitized config template as xdp_tts_service/tts_config.example.json.
• Uses a standalone dependency list for the wheel instead of the full monorepo environment.
• Builds a wheel with CLI entrypoints:
  • xdp-tts-service
  • xdp-tts-init-config

The temporary staging directory is deleted automatically after build.

Build

cd /data1/skill_dev/xDataPrep/xdp_service/tts_service
bash build_tts_service_wheel.sh

Wheel will be generated in dist/.

Install

python3 -m pip install dist/xdp_tts_service-0.1.1-py3-none-any.whl

Initialize config

xdp-tts-init-config --output ./tts_config.json

Run service

xdp-tts-service --host 127.0.0.1 --port 5002 --config ./tts_config.json

Notes

• This wheel does not include model weights.
• Generated config uses placeholder model paths. Update them before running.
• cosyvoice3_0.5b_openvino is now supported in tts_config.json.
• For cosyvoice3_0.5b_openvino, set model_dir/ov_model_dir to the OpenVINO-exported directory and original_model_dir to the original CosyVoice3 model directory.
• The built wheel bundles the CosyVoice3 OpenVINO helper as text_to_speech/cosyvoice3_ov_helper.py.
• If the CosyVoice helper is not in the default workspace location, set XDP_COSYVOICE3_OV_HELPER=/absolute/path/to/ov_cosyvoice_helper.py.
• If the CosyVoice source repo is not in the default location, set XDP_COSYVOICE_REPO=/absolute/path/to/CosyVoice.
• For qwen3_tts_0.6b_base_openvino, also set checkpoint_path to the original Base PyTorch checkpoint so voice-clone fallback can load the speech tokenizer weights.
• xDataPrep now vendors the Qwen3 OpenVINO helper in text_to_speech/qwen3_ov_helper.py.
• The built wheel bundles that vendored helper, so neither source-tree runs nor installed runs need the original notebook helper path.
• Re-run build_tts_service_wheel.sh after changing text_to_speech/*.py.
• The wheel is intended to be open-source distributable: no workspace-local helper path and no machine-local default model path are bundled anymore.

Model-Specific Guide

• For a user-facing CosyVoice3 OpenVINO setup and API usage guide, see README_COSYVOICE3.md.