Fabio
c36510c21e
2.1 KiB
OmniVoice tests
Voice-clone benchmark scripts and a Python-to-C++ integration test for the
omnivoice family. All scripts resolve the repository root themselves, so they
can be run from any directory.
They expect models/OmniVoice to be installed:
python3 tools/model_manager.py install omnivoice
Build
cmake -S . -B build -DENGINE_ENABLE_CUDA=ON
cmake --build build --parallel --target audiocpp_cli audiocpp_server
Weight-type benchmarks
run_clone_benchmark.sh sweeps single voice-clone requests and
run_clone_batch_benchmark.sh sweeps --batch-text-file runs, both across
every combination of --num-inference-steps (8 to 64),
omnivoice.generator_weight_type, and omnivoice.audio_tokenizer_weight_type.
./tests/omnivoice/run_clone_benchmark.sh
./tests/omnivoice/run_clone_batch_benchmark.sh
Each script runs 375 combinations, so a full sweep takes hours. Edit
STEPS_LIST, WEIGHT_TYPES, and AUDIO_TOKENIZER_WEIGHT_TYPES at the top to
narrow it. WAVs and a full log are written under tests/omnivoice/outputs/,
which is ignored by Git.
Measured results are in
OmniVoice runtime weight-type benchmark.
The headline finding is that runtime omnivoice.audio_tokenizer_weight_type=q8_0
produces unusable audio without being any faster, while
omnivoice.generator_weight_type=f16 is 1.26x faster than the native path.
Assets
assets/ holds the shared Bengali fixtures used by every script here:
ref_audio_01.wav ref_text_01.txt longer reference voice
ref_audio_02.wav ref_text_02.txt reference voice used by default
clone_batch_prompts.txt 10 prompts, one request per line
Python-to-C++ integration
python_cpp_simple/ drives OmniVoice from Python while
keeping inference in C++, by sending HTTP requests to audiocpp_server using
only the Python standard library.
Warm benchmark
omnivoice_warm_bench.cpp and omnivoice_python_warm_bench.py are the
framework warmbench helpers driven by tests/warmbench.py, and are unrelated to
the sweep scripts above.