bigstorm/case-benchmark

CASE Benchmark

Carrier-Agnostic Speaker Embedding Benchmark - A comprehensive benchmark for evaluating speaker verification systems under real-world acoustic carrier conditions.

Overview

CASE Benchmark tests speaker embedding models across 24 protocols covering:

• Clean: Matched enrollment and test conditions
• Codecs (7): GSM, G.711 μ-law/A-law, Opus (6k/12k/24k), MP3
• Microphones (7): Laptop, webcam, phone, headset, conference
• Noise (5): SNR levels from 5dB to 25dB
• Reverb (1): Simulated room acoustics
• Playback chains (3): Combined codec + noise + microphone

Quick Start

from case_benchmark.download import download_benchmark

# Download full benchmark (~3.1GB compressed)
download_benchmark("./benchmark")

# Download only specific conditions
download_benchmark("./benchmark", conditions=["clean", "codec"])

# Download only VoxCeleb1-O dataset
download_benchmark("./benchmark", datasets=["voxceleb1_o"])

Structure

Audio files are stored as compressed archives per condition:

audio/
├── voxceleb1_o_clean.tar.gz     # 69MB  - 400 files
├── voxceleb1_o_codec.tar.gz     # 464MB - 2,800 files
├── voxceleb1_o_mic.tar.gz       # 467MB - 2,800 files
├── voxceleb1_o_noise.tar.gz     # 346MB - 2,000 files
├── voxceleb1_o_reverb.tar.gz    # 71MB  - 400 files
├── voxceleb1_o_playback.tar.gz  # 205MB - 1,200 files
├── librispeech_clean.tar.gz     # 63MB  - 392 files
├── librispeech_codec.tar.gz     # 426MB - 2,744 files
├── librispeech_mic.tar.gz       # 430MB - 2,744 files
├── librispeech_noise.tar.gz     # 331MB - 1,960 files
├── librispeech_reverb.tar.gz    # 67MB  - 392 files
└── librispeech_playback.tar.gz  # 196MB - 1,176 files

trials/
├── clean_clean.txt
├── clean_codec_*.txt    # 7 codec protocols
├── clean_mic_*.txt      # 7 microphone protocols
├── clean_noise_*.txt    # 5 noise protocols
├── clean_reverb.txt
└── clean_playback_*.txt # 3 playback chain protocols

After extraction, audio files are organized as:

voxceleb1_o/
├── clean/{speaker_id}/utt_*.wav
├── codec/{codec_type}/{speaker_id}/utt_*.wav
├── mic/{mic_type}/{speaker_id}/utt_*.wav
├── noise/{snr_level}/{speaker_id}/utt_*.wav
├── reverb/{speaker_id}/utt_*.wav
└── playback/{chain_type}/{speaker_id}/utt_*.wav

Trial Format

Each trial file contains lines in the format:

<label> <enrollment_path> <test_path>

Where:

• label: 1 for same speaker, 0 for different speaker
• enrollment_path: Path to enrollment audio (always clean)
• test_path: Path to test audio (condition-dependent)

Datasets

Dataset	Speakers	Utterances	Source
VoxCeleb1-O	40	400 clean	VoxCeleb1 test set
LibriSpeech	40	392 clean	LibriSpeech test-clean

Leaderboard

Rank	Model	Absolute EER	Degradation	Clean EER
1	WeSpeaker ResNet34	3.01%	+2.43%	0.58%
2	SpeechBrain ECAPA-TDNN	3.05%	+2.49%	0.56%
3	CASE HF v2-512	3.53%	+2.31%	1.22%
4	NeMo TitaNet-L	4.05%	+3.39%	0.66%
5	pyannote Embedding	4.47%	+2.79%	1.68%
6	Resemblyzer	10.49%	+5.65%	4.84%

See full results for detailed per-protocol breakdowns.

Related Resources

Resource	Description	Link
CASE HF v2-512	Carrier-agnostic speaker embedding model	HuggingFace Model
Benchmark Code	Evaluation scripts and tools	GitHub
Metrics Guide	How to interpret Clean EER, Degradation Factor	Metrics Documentation
Submission Guide	How to submit your model to the leaderboard	Submission Guide

License

• Audio data: CC BY-NC 4.0 (non-commercial use)
• Evaluation code: MIT License

Citation

@misc{case-benchmark-2026,
  title={CASE Benchmark: Carrier-Agnostic Speaker Embedding Evaluation},
  author={Gitter, Ben},
  year={2026},
  url={https://github.com/gittb/case-benchmark}
}