AirCaps/speaker-id-noise

AirCaps speaker ID noise benchmark

This repository publishes the exact reconstruction plan, protocol, validation records, evaluation code, checksums, and compact model results for a deterministic VoxCeleb1-O + MUSAN robustness benchmark.

Audio availability

This is a reproducibility package, not an audio mirror. The clean and rendered noisy WAV files cannot be redistributed through this public repository. VoxCeleb licenses its supplied metadata under CC BY-SA 4.0, but that metadata license does not provide a corresponding blanket right to redistribute the underlying YouTube-derived recordings. The official VoxCeleb site also no longer distributes the audio files. Although MUSAN is CC BY 4.0, mixing MUSAN with VoxCeleb does not remove or replace the rights attached to the VoxCeleb source audio.

For that reason, this repository contains identifiers, deterministic construction instructions, validation records, and expected output hashes—but no source or derived audio. Users must obtain VoxCeleb1 and MUSAN separately through channels for which they have appropriate access and rights.

To reconstruct the benchmark, obtain VoxCeleb1 and MUSAN through channels for which you have appropriate access and rights, then use the included deterministic construction plan and scripts. Every rendered waveform has an expected SHA-256 hash in manifests/mixture_manifest.jsonl.

Benchmark contents

• 4,874 clean VoxCeleb1 test utterances.
• 3 MUSAN categories: speech/babble, music, and environmental noise.
• 7 requested SNRs per category: -10, -5, 0, 5, 10, 15, and 20 dB.
• 21 noisy conditions and 102,354 rendered noisy waveforms.
• 16 kHz mono PCM16 output.
• Deterministic seed: 1234.
• Full-utterance cosine speaker-verification scoring.

Speech/babble mixtures use 3–6 MUSAN speech sources; music and environmental-noise mixtures use one source. Noise segments are selected deterministically and extended to the clean utterance duration.

No peak normalization was applied before PCM16 writing, matching the benchmark used for the reported experiments. Consequently, 9.61% of noisy outputs contain at least one clipped sample. Use the measured SNR and clipping fields in the manifest when analyzing results; requested SNR is not always identical to post-write physical SNR.

Protocols

Two VoxCeleb1-O trial lists are preserved:

Protocol	Trials	Positive	Negative	Unique references	SHA-256
Original veri_test.txt	37,720	18,860	18,860	4,715	303b2b657042a27bf465d4c8aa84e12765373cdc01046665241ccd5783bd5976
Cleaned veri_test2.txt	37,611	18,802	18,809	4,708	0bc0a0fe3e557f1a75fb71e566d862d460709e80a4fe28e80e49bc0ab3a536ea

For noisy evaluation, each trial is scored as clean enrollment against noisy probe in both directions and the two cosine scores are averaged. Pooled EER concatenates scores and labels across all included conditions before calculating EER.

Headline cleaned-protocol results

EER in percent; lower is better.

Model	Clean	All noisy	Babble	Music	Noise
A5sV2-spk-id	0.612	5.375	9.339	3.099	2.279
SimAM-ResNet293 pretrained	0.707	8.204	16.709	3.769	3.040
CAM++-LM English	0.707	7.721	15.243	3.889	3.136
ECAPA-TDNN English	0.862	8.507	15.724	5.288	3.166
TitaNet-Large	0.824	8.825	16.575	5.558	2.928
ERes2NetV2	6.385	15.688	20.849	14.309	10.834
WavLM Base+ SV	4.716	15.779	20.672	13.619	13.050

The curated machine-readable results are under results/. Only the pretrained SimAM checkpoint is included; ERes2NetV2 LMFT and W2V-BERT 2.0-LM results are intentionally excluded.

Evaluated checkpoint provenance

Every model name in the result table links to the exact checkpoint file used for evaluation. Public checkpoints are pinned to an immutable repository revision or a versioned ModelScope release. SHA-256 values below are hashes of the evaluated local files, not just model-repository identifiers.

Display name	Exact evaluated artifact	SHA-256
A5sV2-spk-id	Private model	—
SimAM-ResNet293 pretrained	Zyphra/Zonos-v0.1-speaker-embedding@9fe3dbd → ResNet293_SimAM_ASP_base.pt	9f02677cd204f91e228f9ec344b7262749a9ab80b9848a7e0452e198d629a65d
CAM++-LM English	Wespeaker/wespeaker-voxceleb-campplus-LM@c5e01c6 → avg_model.pt	3ced11d16c68ee4d1b9b8aea2494f56516d7cc07782d451fb437480ad00a6ac7
ECAPA-TDNN English	iic/speech_ecapa-tdnn_sv_en_voxceleb_16k@v1.0.1 → ecapa_tdnn.bin	e7f8e824f57fc4bc349239963f52143c4a09b825df82296100559f3120653ec1
TitaNet-Large	nvidia/speakerverification_en_titanet_large@0dc382f → speakerverification_en_titanet_large.nemo	e838520693f269e7984f55bc8eb3c2d60ccf246bf4b896d4be9bcabe3e4b0fe3
ERes2NetV2	iic/speech_eres2netv2_sv_zh-cn_16k-common@v1.0.1 → pretrained_eres2netv2.ckpt	0eb4057106b2573dd7b132cf0c36273ab29afd192c1610f80baa9c556dbb963c
WavLM Base+ SV	microsoft/wavlm-base-plus-sv@feb593a → pytorch_model.bin	e906bce2fa42fb497a1d1a9ecf81548adb7e03b12a5644e32d2f42f0d6500fad

Repository layout

manifests/
  construction_plan.jsonl     # deterministic source selections and offsets
  mixture_manifest.jsonl      # measured levels, clipping data and output hashes
  validation_report.json      # dataset-level integrity and SNR audit
protocols/
  veri_test.txt                # original VoxCeleb1-O trials
  veri_test2.txt               # cleaned VoxCeleb1-O trials
code/
  prepare_openslr_musan_exunet_bank.py
  generate_vox1_musan_exunet.py
  vox1_musan_eer.py
results/
  benchmark_summary.csv       # curated model comparison
  benchmark_summary.json      # protocol metadata and exact values
assets/
  speaker-verification-eer.png

Reconstruction

The scripts expose their full command-line arguments with --help. A reconstruction requires:

1. A local VoxCeleb1 test tree whose relative paths match the trial files.
2. A local MUSAN tree from OpenSLR SLR17.
3. A 16 kHz-compatible Python environment with PyTorch, NumPy, SciPy, SoundFile, and tqdm.
4. The supplied construction plan and seed 1234.

After generation, compare the output count, aggregate validation values, and per-file SHA-256 values against the supplied validation records.

Licensing and citation

• The VoxCeleb authors license the supplied metadata under CC BY-SA 4.0. Rights to the underlying source recordings may differ.
• MUSAN is distributed under CC BY 4.0.
• The code in this repository is provided under the MIT License.
• The manifests describe derived experimental metadata; they do not grant rights to source audio.

Please cite VoxCeleb and MUSAN when using this reconstruction package:

@inproceedings{nagrani2017voxceleb,
  title={VoxCeleb: a large-scale speaker identification dataset},
  author={Nagrani, Arsha and Chung, Joon Son and Zisserman, Andrew},
  booktitle={INTERSPEECH},
  year={2017}
}

@misc{snyder2015musan,
  title={MUSAN: A Music, Speech, and Noise Corpus},
  author={Snyder, David and Chen, Guoguo and Povey, Daniel},
  year={2015},
  eprint={1510.08484},
  archivePrefix={arXiv}
}

Sources: VoxCeleb1 and MUSAN/OpenSLR SLR17.