eDIAR: training-free open-speaker streaming diarization for long meetings and large archives

Training-free. This repository contains code only and stores no weights. It lifts the speaker limit of a fixed-slot diarizer (Nemotron-3 Diarization, 8 slots) by clustering speaker embeddings (WeSpeaker ResNet34-LM) inside a slot, and only after every slot has been used. The clustering engine is the official e-means algorithm. The streaming system re-clusters a bounded 90 s window every chunk, may revise labels for 10 s, and anchors permanent speaker ids to the slots of the head.

Demo: https://kyudan1.github.io/eDIAR/

Usage: streaming or non-streaming

One class, one switch. Everything else (models, clustering, speaker ids) is the same.

	streaming=True (default)	streaming=False
Audio	arrives in chunks, causal	whole recording at once
Latency	1.04 s (latency="low_latency") or 0.64 s ("very_low_latency")	none, runs after the recording
Labels	provisional, revised for 10 s, then fixed	final
Clustering	bounded 90 s window, re-run every chunk	all embeddings of the session at once
Use it for	live meetings, long recordings, bounded cost	files, highest accuracy
Paper setting	streaming tables	non-streaming tables

import soundfile as sf
from emeans_diar import SlotEMeansDiarizer

wav, sr = sf.read("meeting.wav", dtype="float32")             # mono

# 1) streaming (default): audio is fed in 100 ms chunks like a live system
dia = SlotEMeansDiarizer(device="cuda", streaming=True, latency="low_latency")
for start, end, speaker in dia.diarize(wav, sr):
    print(f"{start:7.2f} {end:7.2f} spk{speaker}")

# 2) non-streaming: the whole file at once
dia = SlotEMeansDiarizer(device="cuda", streaming=False)
segments = dia.diarize(wav, sr)

# 3) real streaming loop (live audio): incremental interface
dia = SlotEMeansDiarizer(streaming=True)
session = dia.new_session()
for pcm in microphone_chunks():                               # 16 kHz int16 bytes, any chunk size
    msgs = session.feed(pcm)                                  # provisional + revised labels
session.feed(None, flush=True)

Command line, RTTM on stdout:

python examples/diarize.py meeting.wav                        # streaming, 1.04 s
python examples/diarize.py meeting.wav --latency very_low_latency
python examples/diarize.py meeting.wav --offline              # non-streaming

The two upstream models are downloaded from their own repositories on first use (HF_TOKEN if they are gated for you). Install with pip install -r requirements.txt.

Results (streaming, 1.04 s input buffer, DER %, collar 0.25 s, overlap scored)

System	VoxConverse	AMI	ICSI	LibriSpeech-OpenSpk
Streaming Sortformer 4spk v2.1	14.1	8.3	27.7	41.4
Nemotron-3 head (8 slots)	6.4	9.3	27.4	19.8
This repository (after revision)	5.8	9.1	27.4	16.4

Numbers are generated from the result files of the accompanying paper (cropped sessions of 7-8 minutes; 44 / 16 / 15 / 18 sessions).

What this gives you:

• No cap on the number of speakers. The 8-slot Nemotron-3 head becomes open-speaker without retraining; the paper shows the same for the 4-slot Streaming Sortformer.
• Robust on long meetings. Bounded-window re-clustering and slot-anchored ids keep the error flat over hour-long meetings.
• Linking speakers across recordings. Frame-level linking with e-means finds the same person across an archive of more than a thousand speakers, reaching the loss of Lloyd's k-means several times faster.
• Low and constant compute. The cost per update does not grow with the meeting (real-time factor about 0.09 on a shared CPU host).

Licenses

Code: Apache-2.0 (see LICENSE, NOTICE). Vendored e-means: MIT (Meta). The models it downloads keep their own licenses (Nemotron-3 Diarization: openmdw-1.1; WeSpeaker ResNet34-LM: see its model card).