voice-detect.cpp - Parity report

This document records the numerical and end-to-end parity of the C++/ggml speaker-recognition path against the reference implementations (SpeechBrain for ECAPA-TDNN; the ONNX runtime graph for WeSpeaker / 3D-Speaker encoders). All comparisons are CPU, deterministic (dither off).

Status: scaffolding. The encoder graphs are not implemented yet, so the matrix below is the target each model must hit before it is marked PASS.

Parity gates

A stage or model passes when it meets ALL of:

1. Embedding cosine similarity >= 0.9999 vs the reference L2-normalized speaker embedding for the same clip.
2. Max absolute difference <= 1e-3 on every dumped intermediate tensor (fbank, encoder_out, embedding) vs scripts/gen_baseline.py.
3. Identical verification verdict vs the reference on a fixed enroll/probe set: for every (clip_a, clip_b, threshold) triple the C++ verified flag matches the reference decision (same-speaker / different-speaker).

The per-stage golden compare is tests/parity.hpp (vdtest::compare, vdtest::cosine); tests/test_fbank.cpp is the first gate. Tests SKIP (exit 77) when the baseline GGUF env vars are unset, so CI without the reference venv never fails on them.

Model coverage matrix (target)

Checkpoint	Family	arch	embedding	FBank	Embedding cosine	Verdict	Status
speechbrain/spkrec-ecapa-voxceleb	ECAPA-TDNN	ecapa_tdnn	192	80	-	-	TODO
WeSpeaker ResNet34	ResNet34	wespeaker_resnet34	256	80	-	-	TODO
3D-Speaker ERes2Net (base 200k)	ERes2Net	eres2net	512	80	1.000000	-	GREEN
3D-Speaker CAM++ (zh-cn 16k-common)	CAM++	campplus	192	80	1.000000	-	GREEN
superb/wav2vec2-base-superb-er	wav2vec2 emotion	wav2vec2_emotion	-	-	emotion_probs max|d| 4.8e-7	dominant matches	GREEN
audeering/wav2vec2-large-robust-24-ft-age-gender	wav2vec2 age/gender	wav2vec2_age_gender	-	-	age 72.12==72.12; gender_probs max|d| 6e-8	dominant matches	GREEN

Stage parity (target)

Stage	Reference	Tensor	Gate
FBank front end	torchaudio.compliance.kaldi.fbank + CMN	fbank [n_mels, T]	max|d| <= 1e-3
Speaker encoder	SpeechBrain / ONNX forward	encoder_out [d, T']	max|d| <= 1e-3
Embedding	L2-normalized speaker vector	embedding [embedding_dim]	cosine >= 0.9999

The FBank stage is the load-bearing one: any deviation from the reference Kaldi options (POVEY window, 0.97 pre-emphasis, dither off, snip_edges, mel band edges, log floor, per-utterance CMN) propagates into a verification-threshold drift. See src/fbank.cpp.