How narcotic-sh/senko works
Project type and purpose
• Senko is a high-performance speaker diarization pipeline written in Python, with a C++ submodule for feature extraction.
• It implements a four-stage data-processing pipeline: VAD → Fbank feature extraction → speaker embedding inference → clustering → output formatting.
File structure
• Top level:
• senko/ – the core Python package containing orchestrator (diarizer.py), modules (camplusplus.py, cluster/, fbank_extractor/), and utilities.
• models/ – pre-trained model artifacts and configs.
• examples/, evaluation/, tracing/ – scripts for demonstration, benchmarking, and tracing/optimizations.
• pyproject.toml, CMakeLists.txt – build/config specifications for Python packaging and C++ fbank extraction.
• Inside senko/fbank_extractor/: a CMake-based C++ library for mel-filterbank extraction.
• Inside senko/cluster/: two implementations (CPU and GPU via RAPIDS cuML), with YAML configs under conf/.
README for insights
• Confirms the four-stage pipeline and external dependencies (Pyannote or Silero for VAD, PyTorch for embeddings, RAPIDS for GPU clustering).
• Describes use cases, performance targets, and guidance on GPU vs CPU execution.