Overview of voice-classifier-cpp
developThe voice-classifier-cpp package is a standalone C library and a set of GGUF conversion scripts. It ports native voice-side classifiers to the elizaOS/llama.cpp fork's ggml dispatcher. This is intended to replace the current onnxruntime-node path used by voice services in plugins/plugin-local-inference/src/services/voice/.
Supported classifiers include:
- Voice emotion classifier: Provides 7-class basic-emotion soft probabilities.
- End-of-turn (EOT) detector: Returns audio-side $P(\text{end_of_turn}) \in [0, 1]$. Note: Currently returns
-ENOSYSuntil an upstream audio-turn model is pinned. - Speaker embedding encoder: Provides 256-dim WeSpeaker embeddings with a cosine-distance helper.
- Diarizer: Uses pyannote-3 segmentation powerset labels.
Important: Before modifying this directory, read AGENTS.md for the full port plan, including upstream pins, GGUF schemas, and integration steps.