Qwen3-TTS OpenVINO IR
This repository stores exported OpenVINO IR artifacts for qwen3-tts-openvino runtime downloads.
Layout
openvino_realtime/voice_design/: voice_design OpenVINO IRopenvino_realtime/custom_voice/: custom_voice OpenVINO IRopenvino_realtime/base/: VoiceClone/Base OpenVINO IR
The runtime sidecar downloads these files automatically when a mode is missing locally.
Profile
Current public profile: runtime-minimal.
For runtime-minimal, each mode keeps only the validated low-memory production graph set:
- text and codec embedding graphs
int8_sym_batch_fused_gqa paged-KV batch talker seed graph- cached standalone subcode graph, executed row-by-row by the online scheduler
- streaming decoders
c0_t8, c25_t12, and c25_t24
- VoiceClone/Base additionally keeps
speech_encoder, speaker_encoder, and code_frame_embedding
Legacy full decoders, alternate streaming chunks, no-cache talker graphs, batch-fused decode graphs, sampled batch subcode graphs, fixed-bucket/unroll diagnostic graphs, and OpenVINO compile caches are intentionally not published here.
