rewayai/suplime-large

SUPlime-L — speaker diarization (WavLM-Large)

SUP as in "what's up": who is speaking, and when.

Open-weights speaker diarization by Re:WayAI, packaged for pyannote.audio 4.x. One 16 kHz mono recording in, who-spoke-when out; no speaker count needed.

This is the WavLM-Large model: 349 M parameters, and the leader of pyannote's 8-corpus benchmark — 15.86 macro DER, ahead of pyannoteAI's commercial precision-2 (16.06). Its sibling SUPlime runs a WavLM-Base+ backbone (114 M) and is the better model across all 12 corpora we score. Same recipe, same embedding, same clustering, same threshold; pick by domain, see Which variant.

Weights are released under CC BY-NC 4.0 (non-commercial) because part of the training data is licensed for research use only, which rules out commercial use of models derived from it — see LICENSE and the rationale below.

Quick start

pip install suplime      # installs pyannote.audio>=4.0.7 and the model code

pip install does not bring FFmpeg, and pyannote.audio reads audio files through torchcodec, which loads FFmpeg's shared libraries at runtime — without them the first call raises RuntimeError: Could not load libtorchcodec. Install it into the same environment:

micromamba install -c conda-forge ffmpeg     # or conda/mamba; on Linux also apt install ffmpeg

Passing a waveform instead of a path needs no decoder at all:

import soundfile as sf, torch
x, sr = sf.read("meeting.wav", dtype="float32", always_2d=True)
output = pipeline({"waveform": torch.from_numpy(x.T), "sample_rate": sr})   # (channel, time)

import torch
from pyannote.audio import Pipeline

pipeline = Pipeline.from_pretrained("rewayai/suplime-large").to(torch.device("cuda"))
output = pipeline("meeting.wav")
for turn, _, speaker in output.speaker_diarization.itertracks(yield_label=True):
    print(f"{turn.start:.2f} {turn.end:.2f} {speaker}")

No Hugging Face token or gated access is required. The suplime package is needed because the checkpoints reference model classes that do not exist in pyannote.audio itself; without it pyannote raises MissingDependency.

Optional knobs:

pipeline("meeting.wav", num_speakers=3)          # or min_speakers= / max_speakers=
pipeline.instantiate({"clustering": {"threshold": 0.68, "min_cluster_size": 12, "method": "centroid"},
                      "segmentation": {"min_duration_off": 0.0}})   # lower threshold = more speakers

SUPLIME_FP16=0 runs the WavLM backbone in fp32 (default: fp16 autocast on CUDA during inference — all numbers below were produced with the default).

Architecture

Stage	Model	Details
Segmentation	WavLM-Large → Conformer → powerset	24 WavLM layers mixed with learned softmax weights; 4 Conformer layers (4 heads, FFN 256, depthwise kernel 31, dropout 0.1); 2 linear layers (128); powerset output over 4 speakers with ≤ 2 simultaneous (11 classes); 10 s windows, 20 ms frames, window step 1 s. 349 M parameters.
Embedding	WeSpeaker SimAM-ResNet34 + attentive statistics pooling	voxblink2_samresnet34_ft (VoxBlink2 pre-training, VoxCeleb2 fine-tuning), converted to a pyannote model; 256-dim, cosine metric; overlapping speech excluded from pooling. 25 M parameters, shared with SUPlime byte for byte.
Clustering	Agglomerative, centroid linkage	threshold 0.72, min_cluster_size 12 (pyannote's standard AHC).

The segmentation model was fine-tuned end to end (WavLM unfrozen) on the training splits of 11 public corpora — AMI (IHM and SDM), ICSI, VoxConverse, CHiME-6, DiPCo, NOTSOFAR-1, AISHELL-4, AliMeeting, MSDWild, RAMC and AVA-AVD (4,187 files) — with in-batch speaker mixing, simulated-RIR reverberation and MUSAN noise; AdamW, lr 5e-5, weight decay 0.01, OneCycle schedule, gradient clipping 0.5, bf16 mixed precision, 10 s chunks. Training ran in two stages: 15 epochs at effective batch 32 on one GPU, then 20 epochs at effective batch 192 across 8 GPUs under a fresh OneCycle (same peak lr), with early stopping (patience 10). The released weights are the average of the 5 best validation checkpoints. The complete recipe — training splits, augmentation data, one suplime-train command on stock pyannote.audio — is published in training/ of the GitHub repository — pass --wavlm WAVLM_LARGE for this variant. Repository layout follows pyannote's conventions: config.yaml (pipeline), segmentation/pytorch_model.bin, embedding/pytorch_model.bin.

Results

Diarization error rate (%), lower is better:

Corpus (test unless noted)	SUPlime-L	SUPlime	DiariZen-L-s80-v2	pyannoteAI precision-2	pyannote community-1
AISHELL-4	11.21	11.55	10.1	11.4	11.7
AliMeeting (far, ch1)	14.55	14.45	10.8	15.2	20.3
AMI (IHM, Mix-Headset)	11.91	12.59	25.69 †	12.9	17.0
AMI (SDM)	14.85	15.14	13.9	15.6	19.9
AVA-AVD	36.83	38.09	42.62 †	37.1	44.6
MSDWild (few.val)	17.48	17.81	15.8	17.3	22.8
RAMC	11.10	10.86	11.0	10.5	20.8
VoxConverse (v0.3)	8.93	9.21	9.1	8.5	11.2
macro average (8)	15.86	16.21	17.38	16.06	21.04
NOTSOFAR-1 (80-session split)	22.70	20.04	18.86 †	—	27.67 †
ICSI	22.77	22.97	26.54 †	—	30.84 †
CHiME-6	48.98	48.11	48.39 †	—	51.98 †
DiPCo	30.92	30.44	37.56 †	—	34.38 †
macro average (12)	21.02	20.94	22.53	—	26.10

SUPlime-L leads the 8-corpus benchmark at 15.86 macro DER — ahead of precision-2 (16.06), SUPlime (16.21) and DiariZen-L-s80-v2 (17.38) — while across all 12 corpora SUPlime takes it back by 0.06, the four extra sets being far-field and dinner-party audio where the larger backbone does not pay off. Both beat DiariZen by ~1.5 DER on the 12-corpus average, with the largest margins on close-talk AMI and on CHiME-6 / DiPCo, while DiariZen stays clearly better on the Mandarin meeting corpora and MSDWild. The per-corpus hypothesis RTTMs behind every number are in reproducible_research/.

Scoring conditions. Collar 0 s, overlapped speech scored, no oracle speaker count, reference cropped to each corpus' UEM, one clustering threshold (0.72) everywhere. The other systems' columns are their authors' published numbers (DiariZen, pyannote), except where † marks our own re-run of their open-source pipeline, unchanged and on the same files — and for NOTSOFAR-1 the DiariZen authors report 16.7 on a different session split.

Which variant

	SUPlime-L (this model)	SUPlime
Backbone	WavLM-Large	WavLM-Base+
Parameters (segmentation)	349 M	114 M
macro DER, 8-corpus benchmark	15.86	16.21
macro DER, all 12 corpora	21.02	20.94

Take SUPlime-L for meeting and conversational audio of the kind the 8-corpus benchmark covers. Take SUPlime for far-field and dinner-party audio (CHiME-6, DiPCo, NOTSOFAR-1), or when you want the smaller model — it stays within 0.4 DER of the larger one almost everywhere. They share the embedding, the clustering and the operating threshold, so switching is a one-line change.

Limitations

• Threshold. 0.72 is a DER-optimal compromise across corpora, and the spread around it is wide for this model. AISHELL-4 and AliMeeting want 0.68 (10.57 and 13.61) and collapse above 0.74 (AISHELL-4: 20.80 at 0.80); CHiME-6, AMI-SDM and VoxConverse want 0.80, and CHiME-6 alone gains 5.7 DER there (49.14 → 43.44). Tune on your own data — on this model it is worth more than on the base one. A DER-optimal threshold tends to over-merge speakers; if your downstream metric is speaker-attributed WER, a slightly higher threshold is usually better.
• Ahead of SUPlime by 0.35 DER on the 8-corpus benchmark, but behind it on the 12-corpus macro average.
• Results depend on segmentation_batch_size. WavLM-Large is used through torchaudio's normalising bundle wrapper, which layer-normalises over the whole input tensor — so a chunk's scores depend on the other chunks batched with it, and a file's zero-padded last chunk is normalised against its own padding. Measured: scoring four 10 s chunks together rather than singly moves scores by up to 2.8 in log-space. Every number here was produced with the shipped segmentation_batch_size: 32; change it and you get slightly different output. This matches torchaudio's own *_LARGE bundles (pipelines/_wav2vec2/utils.py), so it is shared with every system built on them — SUPlime (Base+) is unaffected, its bundle does not normalise.
• The powerset head models at most 2 simultaneous speakers per frame and 4 speakers per 10 s window; recordings with heavy 3-way overlap are under-served.
• Trained on 16 kHz meeting / conversational / broadcast / movie audio; telephone (8 kHz) speech is out of domain and is resampled, not modelled.
• The embedding model is trained on VoxBlink2 / VoxCeleb2 (largely English YouTube speech); no fairness evaluation across languages or demographics has been done.
• English and Mandarin dominate the training data.

License and attribution

Model weights (segmentation/, embedding/): CC BY-NC 4.0. The segmentation model was trained on corpora with mixed terms: several are licensed for research / non-commercial use only (RAMC, MSDWild, AVA-AVD) and others are share-alike (AISHELL-4, AliMeeting, CC BY-SA 4.0). A model derived from that data is bound by the most restrictive of those terms, so the weights cannot be offered for commercial use and are released under CC BY-NC 4.0. The embedding weights are a conversion of WeSpeaker's voxblink2_samresnet34_ft and remain subject to WeSpeaker's and the VoxBlink2 / VoxCeleb2 data terms.

Code (suplime package): MIT; the SimAM-ResNet backbone is Apache-2.0, ported from WeSpeaker.

Built on pyannote.audio (Bredin et al.), WavLM (Chen et al., Microsoft, MIT) and WeSpeaker (Wang et al.). The powerset segmentation formulation follows Plaquet & Bredin, Powerset multi-class cross entropy loss for neural speaker diarization (Interspeech 2023); the WavLM + Conformer segmentation design is in the spirit of DiariZen (Han et al., BUT).

Citation

@misc{suplimel2026,
  title  = {SUPlime-L: open-weights speaker diarization for pyannote.audio},
  author = {Re:WayAI},
  year   = {2026},
  url    = {https://huggingface.co/rewayai/suplime-large}
}

Diarization is not enough?

SUPlime tells you who spoke and when. If you also want to know what they said, and who they actually are, that is our day job: Re:WayAI is the speech API that knows who's talking.

• Transcription in 25 European languages (40 for real-time streaming), with overlap-robust diarization from the same team that built SUPlime
• Voice enrollment: speakers get their real names, not SPEAKER_03
• Real-time streaming over WebSocket, transcript Q&A, one-click DOCX / PDF exports
• On-premise deployment, audio deleted right after processing, ISO 27001

Pay-as-you-go, 50 free hours to start, no subscription. The lime approves.