voice-classifier-cpp

Standalone C library + GGUF conversion scripts that port three small voice-side classifiers to the elizaOS/llama.cpp fork's ggml dispatcher, replacing the onnxruntime-node path used today by the voice services in plugins/plugin-local-inference/src/services/voice/:

Voice emotion classifier — 7-class basic-emotion soft probabilities.
End-of-turn detector — audio-side P(end_of_turn) ∈ [0, 1].
Speaker embedding encoder — 192-dim ECAPA-style embedding plus a cosine-distance helper.

Today the model entry points are stubs (every voice_*_open / voice_*_classify / voice_*_score / voice_*_embed returns -ENOSYS). The shared utilities are real:

voice_emotion_class_name returns the canonical class names in the pinned 7-class order;
voice_speaker_distance is a real cosine-distance implementation;
voice_mel_compute is a real shared log-mel front-end (n_mels=80, n_fft=512, hop=160 at 16 kHz).

The full port plan — upstream pins, GGUF schema per head, fork integration steps, replacement path for the TS services — lives in AGENTS.md. Read that before changing anything in this directory.

Build

cmake -B packages/native/plugins/voice-classifier-cpp/build \
      -S packages/native/plugins/voice-classifier-cpp
cmake --build packages/native/plugins/voice-classifier-cpp/build -j
ctest --test-dir packages/native/plugins/voice-classifier-cpp/build --output-on-failure

Builds libvoice_classifier.a and four test binaries:

Test	What it asserts
`voice_classifier_stub_smoke`	The three model heads still return `-ENOSYS` and clear out-args
`voice_emotion_classes_test`	The 7-class vocabulary order is intact
`voice_speaker_distance_test`	Cosine distance: identical=0, orthogonal=1, opposite=2
`voice_mel_features_test`	A 1 kHz sine wave peaks in the low-mid mel band

Layout

include/voice_classifier/voice_classifier.h  Public C ABI (frozen — see AGENTS.md).
src/voice_classifier_stub.c                  ENOSYS stub for the three model heads.
src/voice_emotion_classes.c                  Real: 7-class name table.
src/voice_speaker_distance.c                 Real: cosine distance helper.
src/voice_mel_features.c                     Real: shared log-mel front-end.
scripts/voice_emotion_to_gguf.py             Skeleton converter (TODO blocks documented).
scripts/voice_eot_to_gguf.py                 Skeleton converter.
scripts/voice_speaker_to_gguf.py             Skeleton converter.
test/                                        Four ctest binaries; see table above.
CMakeLists.txt                               Builds libvoice_classifier + tests.

License

Apache 2.0 — matches the suggested ECAPA-TDNN upstream (speechbrain/spkrec-ecapa-voxceleb). The pinned upstream commits for each head are recorded in the corresponding scripts/voice_*_to_gguf.py file at conversion time.