ACloudCenter/voice-attribution-proto

Self-Hosted Voice-Attribution Prototype

Offline speaker-attribution pipeline for white-label meeting platforms: enrollment gallery + offline diarization + voiceprint (ECAPA) cluster matching, measured on labeled multi-speaker audio. All models are self-hostable and commercial-safe (CCBY-4.0 / Apache-2.0 / OpenMDW-1.1).

Files

File	What it does
attribution_proto.py	Shared core: LibriSpeech synthetic-meeting builder, ECAPA enrollment, gallery matching. selftest stage (CPU) validated the matcher: intra-speaker cosine 0.79-0.86 vs cross-speaker 0.16-0.37.
attribution_nemo.py	GPU stage: NeMo diar_sortformer_4spk-v1 chunked into 80s windows (model session cap 90s), ECAPA cluster merge across windows, gallery naming.
attribution_ami.py	Real-meeting eval: AMI ES2004a.Mix-Headset + official manual annotations (segments -> RTTM), enrollment from 3 in-meeting segments per speaker. Diarizer is swappable via bootstrap patches (bootstrap_community1.py, bootstrap_nemotron3.py).
bootstrap_community1.py	Run the eval with pyannote speaker-diarization-community-1 (gated; ffmpeg + hub<1.0 + weights_only patch required).
bootstrap_nemotron3.py	Run the eval with nvidia/Nemotron-3-Diarization via the model card's Transformers API (public, ungated, offline config).
results_sortformer_chunked.json	Synthetic 4-spk meeting results.
results_ami_es2004a.json	Real AMI meeting results (Sortformer chunked).
results_ami_es2004a_community1.json	Real AMI meeting results (pyannote community-1).
results_ami_es2004a_nemotron3.json	Real AMI meeting results (Nemotron-3-Diarization).

Results (AMI ES2004a.Mix-Headset — all rows: same recording, same ECAPA gallery, same scorer, manual-annotation references)

Diarizer	Attribution accuracy	DER (manual refs)	Diarization RTF
Sortformer v1 chunked (80s windows)	67.9-70.7%	37.4%	0.026
pyannote community-1	70.5%	24.9%	0.023
Nemotron-3-Diarization (2026-09-23)	71.9%	34.3%†	0.0034

† Under manual segment references. NOT comparable to the 9-11% on NVIDIA's model card, which uses forced-alignment references. Under manual references, tight frame-level boundaries are penalized (reference labels include within-segment silence as speech), so this row cannot be read as "worse diarizer than community-1".

Synthetic 4-spk meeting (7 min, 10% overlap): Sortformer chunked = 91.7% attribution, 17.3% DER, RTF 0.022.

Key observations

• Enrollment matching is solved and stable: cluster-to-name cosine scores were 0.85-0.95 in every run across three diarizers (synthetic and real audio), with 3 clips per speaker. All diarizers produced 4 clusters; every cluster was named one-to-one, none misnamed.
• Attribution accuracy plateaus at ~70-72% across three diarizer generations (DER under one fixed protocol: 37.4% -> 24.9% -> 34.3%): named attribution does not track DER. The residual error is overlap/crosstalk regions where the single mixed mic contains blended speech and single-label references disagree by construction. In deployment, per-device audio tracks bypass this entirely.
• DER itself is protocol-dependent: the same Nemotron-3 checkpoint scores ~9-11 on NVIDIA's card (forced-alignment refs) vs 34.3 here (manual refs). For the attribution metric, DER under either reference is the wrong yardstick — this is the benchmark-paper thesis.
• Diarizer recommendation: ship nvidia/Nemotron-3-Diarization — 8 speakers (Sortformer v1 capped at 4), no session-length cap, ~6x faster than community-1 (RTF 0.0034), public/ungated, commercial license (OpenMDW-1.1), native Transformers support, and streaming configs down to 0.32s latency for live in-meeting labeling. pyannote community-1 remains the validated fallback. Nemotron-3 must still be paired with enrollment naming — its output channels are arrival-order anonymous speakers, exactly what the ECAPA gallery matcher names.

Disclosure

In the AMI eval, enrollment clips are reference segments from the same recording. This simulates enrollment captured via per-device tracks or the confirm-speaker UI during the meeting — the flows the product actually uses — and is not cross-corpus enrollment.

How to run

# CPU selftest (matcher validation)
python attribution_proto.py selftest --device cpu

# Synthetic GPU test
python attribution_nemo.py   # stage gputest

# Real AMI meeting (Sortformer)
python attribution_ami.py
# Real AMI meeting (pyannote community-1, gated: needs HF_TOKEN with repo access)
python bootstrap_community1.py   # patches the diarize block, then runs
# Real AMI meeting (Nemotron-3-Diarization, ungated)
python bootstrap_nemotron3.py

Dependencies:

• Sortformer path: nemo_toolkit[asr], speechbrain, datasets, librosa, soundfile, pyannote.database+pyannote.metrics (DER). GPU: any >=8GB.
• community-1 path: pyannote.audio==4.0.0 + ffmpeg (torchcodec) + huggingface_hub<1.0 + torch.load(weights_only=False) patch. Gated models.
• Nemotron-3 path: ffmpeg, transformers (source; needs setuptools for the git build), accelerate, speechbrain, pyannote.database + pyannote.metrics (standalone — do NOT add pyannote.audio, it forces huggingface_hub<1.0 which conflicts with transformers 5.x).