Intel TTS Benchmark

• Hugging Face source repo: hexgrad/Kokoro-82M
• Hugging Face artifact repo: mweinbach1/ai-pc-benchmarks-tts-intel-openvino
• Provider-local staged payloads: model/
• Provider-local vendored runtime: runtime/kokoro/
• Shared prompt: ../assets/test_prompt.txt

• results/TTS/Intel/test_prompt.npu.result.json
• results/TTS/Intel/test_prompt.npu.transcript.txt
• results/TTS/Intel/test_prompt.npu.wav
• results/TTS/Intel/test_prompt.npu.run.log

• The default run path stages config.json, the Kokoro weights, requested voice files, and pre-exported OpenVINO IRs from mweinbach1/ai-pc-benchmarks-tts-intel-openvino.
• If the artifact repo is incomplete or unavailable, the benchmark falls back to hexgrad/Kokoro-82M for source assets and regenerates any missing OpenVINO exports locally.
• The benchmark vendors the Kokoro Python runtime under runtime/kokoro/, so benchmark execution does not depend on the installed kokoro wheel or an editable upstream clone.
• The benchmark keeps the staged source payload under model/, while the default downloaded or regenerated OpenVINO IRs and NPU static variants live under the machine-local cache root in %LOCALAPPDATA%\AI-PC-Benchmarks\model-artifacts\TTS\Intel\kokoro-openvino\.
• NPU runs use split staged OpenVINO models under npu_static/staged/ plus a CPU torch decoder. Those staged IRs are regenerated automatically when the stage export versions change.
• If you want the provider-local tree to hold the generated IRs instead of the machine-local cache, pass --compiled-model-dir workloads\TTS\Intel\model.
• workloads/TTS/Intel/.cache/npuw/ is machine-local compiler cache. It can be regenerated on each machine and does not need to be copied.
• This benchmark is intended to validate the NPU path and does not treat CPU fallback as success.
• This machine was verified with openvino==2026.1.0rc2 from the official OpenVINO nightly wheel index. The benchmark relies on that 2026.1 line for the current NPU-correct path.
• requirements.txt installs the supporting runtime dependencies plus the en_core_web_sm spaCy model explicitly so Misaki does not need to fetch it during the benchmark run.
• The default voice is af_heart and the default language code is a (American English).
• The transcript output stores the synthesized prompt text so the suite can keep a consistent text sidecar alongside the JSON and WAV outputs.