README.md

metadata

license: cc-by-4.0
task_categories:
  - audio-classification
  - speaker-diarization
language:
  - en
pretty_name: Sortformer Diarization Test Set
size_categories:
  - <1K

Sortformer Diarization Test Set

100 real speech samples extracted from LibriSpeech test-clean for speaker diarization testing and benchmarking with NVIDIA Sortformer 4spk-v2 ONNX models.

Dataset Description

• Samples: 100 audio files (WAV, 16 kHz mono)
• Total size: ~60 MB
• Source: LibriSpeech ASR corpus — test-clean subset
• Speakers: 40 unique speakers from LibriSpeech test-clean
• Purpose: Diarization model evaluation, regression testing, ONNX model validation

Usage with Sortformer ONNX

from huggingface_hub import snapshot_download
import soundfile as sf

# Download the test set
dataset_path = snapshot_download("DimQ1/sortformer-diarization-test-set")

# Load audio
audio, sr = sf.read(f"{dataset_path}/audio/ls_real_000.wav")

Diarization Models

Compatible ONNX models available on HuggingFace:

Model	Size	Speed	Repo
Sortformer FP32	470 MB	16× real-time	DimQ1/sortformer-4spk-v2-onnx-fp32-cpu
Sortformer INT8	129 MB	29× real-time	DimQ1/sortformer-4spk-v2-onnx-int8-cpu
Sortformer INT4	73 MB	33× real-time	DimQ1/sortformer-4spk-v2-onnx-int4-cpu

Ground Truth RTTM

Speaker diarization ground truth annotations are provided in rttm/ directory (NIST RTTM format).

Similar Datasets

For larger-scale diarization training and evaluation:

Dataset	Description	Source
LibriSpeech	1000h English read speech	openslr.org/12
VoxCeleb 1&2	7000+ celebrity speakers	robots.ox.ac.uk/~vgg/data/voxceleb
AMI Corpus	100h meeting recordings	groups.inf.ed.ac.uk/ami/corpus
CALLHOME	Multilingual telephone speech	catalog.ldc.upenn.edu/LDC97S42
DIHARD III	Challenging diarization benchmark	dihardchallenge.github.io/dihard3
MUSAN	Music/speech/noise for augmentation	openslr.org/17

License

Derived from LibriSpeech (CC BY 4.0). See LibriSpeech license for details.