anilkeshwani/speech-integration-eval-sets

Speech-integration eval sets: ASR test sets as discrete speech units

English ASR test sets encoded with the four speech tokenizers of the speech-integration project, plus LibriSpeech dev-clean and dev-other sets with word and phone alignments for unit-representation analyses. The tokenizers are the ones used for the MLS English training data:

Slug	Tokenizer	Frame rate	Codebook
hubert_large_ll60k-layer_22	HuBERT Large (LL-60k) layer 22, k-means 5000 (SpireKMeans)	50 Hz	5000
speechtokenizer-rvq_0	SpeechTokenizer speechtokenizer_hubert_avg, RVQ level 0	50 Hz	1024
mimi-srvq_0	Mimi, split-RVQ level 0 (semantic codebook)	12.5 Hz	2048
focalcodec	FocalCodec 50 Hz, 13-bit BSQ	50 Hz	8192

Training data: anilkeshwani/mls-hubert_large_ll60k-layer_22, anilkeshwani/mls-speechtokenizer-rvq_0, anilkeshwani/mls-mimi-srvq_0 and anilkeshwani/mls-focalcodec. The MLS test split is evaluated from those repos and is not repeated here.

Audio is not included.

Sets

Each set has one config per tokenizer, named <set>.<tokenizer>, for example librispeech-test_clean.hubert_large_ll60k-layer_22. Within a set, the four configs have the same rows in the same order with the same ID, transcript and duration.

Set	Source	Split	Rows	Hours	Units (HuBERT / SpeechTokenizer / Mimi / FocalCodec)
librispeech-test_clean	LibriSpeech test-clean	test	2,620	5.40	970,953 / 973,579 / 244,372 / 970,953
librispeech-test_other	LibriSpeech test-other	test	2,939	5.34	959,587 / 962,539 / 241,740 / 959,587
fleurs_en_us-test	FLEURS en_us test	test	647	1.77	318,748 / 319,395 / 80,074 / 318,748
voxpopuli_en-test	VoxPopuli English test (ESB copy)	test	1,842	4.93	885,586 / 887,465 / 222,561 / 885,586
ami_ihm-test	AMI IHM test (ESB copy)	test	12,643	8.68	1,553,174 / 1,565,861 / 396,223 / 1,553,174
librispeech-dev_clean	LibriSpeech dev-clean (analysis set)	validation	2,703	5.39	968,057 / 970,768 / 243,729 / 968,057
librispeech-dev_other	LibriSpeech dev-other (analysis set)	validation	2,864	5.12	919,980 / 922,856 / 231,793 / 919,980

Excluded sets

• common_voice_en-test (Common Voice 9.0 English test (ESB copy)): Mozilla's Common Voice terms ask that no Common Voice dataset be posted or mirrored, in whole or in part, outside Mozilla Data Collective. The set contains the Common Voice transcripts and units computed from its audio, so it is left out. It was built (16,334 clips, 27.00 h) and can be rebuilt locally with scripts/build_dsu_asr_eval_sets.py.

Columns

Column	Type	Description
ID	string	Utterance ID, unique within a set (see per-set notes)
transcript	string	Reference text copied from the source without modification
speech_tokens	list of int64	Unit sequence at the tokenizer frame rate, raw (consecutive repeats kept)
duration	float64	Duration of the source audio in seconds

Per-set notes

• LibriSpeech test: from openslr/librispeech_asr (revision 71cacbfb7e2354c4226d01e70d77d5fca3d04ba1), field text. ID is the LibriSpeech utterance ID (for example 6930-75918-0000). Transcripts are upper case without punctuation.
• FLEURS en_us test: from google/fleurs (revision 70bb2e84b976b7e960aa89f1c648e09c59f894dd). The FLEURS id is a sentence ID shared by several speakers, so ID is <id>_<wav file stem>. transcript is raw_transcription (cased and punctuated).
• VoxPopuli and AMI IHM: the ESB copies in hf-audio/open-asr-leaderboard (revision b6bdcd0beb34f8975dc659796176d88f43aff502), field text, rows in the source order (longest first). ID is the leaderboard id. VoxPopuli transcripts are lower case with some punctuation. AMI transcripts mix case and keep some punctuation.
• LibriSpeech dev-clean and dev-other (analysis sets, under analysis/): from the OpenSLR tarballs dev-clean.tar.gz and dev-other.tar.gz (MD5 checked), transcripts from the *.trans.txt files, rows sorted by ID.

Text normalization is left to scoring time. The project scores with the Whisper English normalizer.

Other files

• <set>/build_info.json: source repo, revision, config, split and transcript field, row count, hours, longest clip and units per tokenizer.
• alignments/librispeech-dev_{clean,other}.parquet: ID, speaker, words, word_start, word_end, phones, phone_start, phone_end for every utterance of LibriSpeech dev-clean (54,402 words) and dev-other (50,948 words). These are the Montreal Forced Aligner alignments of Lugosch et al. (2019) (Zenodo record 2619474), mapped back onto the lower-cased transcript words: MFA's <unk> words were restored and split clitics merged (quilter + s back to quilter's). Phones are 39 ARPAbet phones without stress plus sil; spn is kept apart.
• alignments/mls-validation.parquet: ID, speaker, words, word_start, word_end for the 3,807 MLS dev utterances, taken from the MMS CTC word alignments of the MLS training repos.
• alignments/mls-train_word_counts.json: whitespace-word counts over the 10,808,037 MLS English train transcripts.
• alignments/alignments_report.json: coverage of the alignment files (every utterance aligned).
• time_maps/<set>/<tokenizer>[-bpe<N>].parquet: per utterance of LibriSpeech dev-clean, dev-other and MLS validation, the units after deduplication (and after speech BPE for the -bpe<N> files), the raw-frame interval each unit covers, word spans in unit positions and, for LibriSpeech, the phone labels. The BPE models are in anilkeshwani/speech-integration-speech-bpe.
• verification/: the JSON reports described under "Verification".
• scripts/: the code that built and checked the sets.

How the sets were built

All audio was decoded with soundfile, downmixed to mono and converted to float32. Audio not at 16 kHz was resampled to 16 kHz with torchaudio.functional.resample (default sinc_interp_hann, width 6), so every tokenizer sees the same signal format as the 16 kHz MLS FLAC training audio. Each utterance was encoded alone, without batching or padding, in fp32.

Slug	Model and revision	Details
hubert_large_ll60k-layer_22	facebook/hubert-large-ll60k @ ff022d095678a2995f3c49bab18a96a9e553f782 with k-means utter-project/SpireKMeans @ bb0d1a04213141f3c61912989abec9641943af60 (kmeans_model, 5000 x 1024)	whole-waveform layer norm, output of transformer layer 22 before the final encoder LayerNorm, nearest centroid by squared L2
speechtokenizer-rvq_0	fnlp/SpeechTokenizer @ 4d54939beef00572fa7bfe41ee35a335c3732f51, speechtokenizer_hubert_avg	speechtokenizer==1.0.1, model.encode(wav[None, None])[0]
mimi-srvq_0	kyutai/mimi @ 89091b3e466eb6a9d11e537bf26b144f194978f7	transformers.MimiModel, 16 kHz to 24 kHz with torchaudio.functional.resample, encode(num_quantizers=8).audio_codes[:, 0]
focalcodec	lucadellalib/focalcodec_50hz @ d6d6d9524e52155c85193c2c3b8da1cf8842f019	focalcodec code at commit 68be6ce, sig_to_toks

For VoxPopuli and AMI, every tokenizer encoded clips longer than 30 s on the CPU to save GPU memory (12 VoxPopuli clips, no AMI clip). CPU and GPU fp32 outputs can differ in a small fraction of units.

The code is in scripts/ (build_dsu_asr_eval_sets.py, build_librispeech_dev_sets.py, dsu_tokenizers.py, check_dsu_asr_eval_sets.py, verify_tokenizers_against_mls.py and requirements-tokenizers.txt, the exact package versions). The alignments and time maps were written by scripts/data/build_analysis_sets.py of the speech-integration project. The sets were stored with datasets.save_to_disk and converted to parquet for this repo; each parquet file was read back and compared equal to its source.

Verification

1. Tokenizers against the MLS training units (verification/mls_verification_flac_100.json): 100 MLS English FLAC files (seed 1, from the first 600 MB of the official MLS English archive) were re-encoded and compared with the stored units of the four MLS training repos. Lengths were equal for every utterance and tokenizer.

   Tokenizer	Utterances identical	Units equal	Worst utterance
   hubert_large_ll60k-layer_22	100 / 100	100%	100%
   speechtokenizer-rvq_0	95 / 100	99.993%	99.86%
   mimi-srvq_0	93 / 100	99.958%	99.15%
   focalcodec	5 / 100	99.45%	98.58%

   HuBERT is bit exact. The SpeechTokenizer and Mimi differences are 1 or 2 units in otherwise identical utterances. Every FocalCodec difference is a flip of one of the 13 BSQ bits at a frame whose latent is near zero, which comes from floating-point noise. verification/mls_verification_opus_20.json repeats the check on Opus re-encodes of MLS (parler-tts/mls_eng) and shows that lossy re-encoding changes 8% to 57% of the units, which is why the FLAC check is the reference.

2. LibriSpeech HuBERT cross-check (verification/eval_sets_check.json): against anilkeshwani/librispeech-hubert_large_ll60k-layer_22 (revision 295b7e5), made earlier with the fairseq pipeline: test-clean 2617 of 2620 utterances identical (99.9997% of units), test-other 2939 of 2939 identical.

3. Post-build checks (verification/eval_sets_check.json), for every set and tokenizer: unique IDs; identical ID, transcript and duration order across tokenizers; every unit inside its codebook; no empty unit sequence; median units per second within 5% of the nominal frame rate.

Loading

from datasets import load_dataset

ds = load_dataset("anilkeshwani/speech-integration-eval-sets", "librispeech-test_clean.mimi-srvq_0", split="test")
print(ds[0]["ID"], ds[0]["transcript"], len(ds[0]["speech_tokens"]))

dev = load_dataset("anilkeshwani/speech-integration-eval-sets", "librispeech-dev_clean.focalcodec", split="validation")

Licenses and attribution

Each set keeps the license of its source. The units and copied transcripts are derived from the source audio and text, so the source terms apply to them. See also LICENSE.md.

Source	License	Attribution	Verified at
LibriSpeech (test-clean, test-other, dev-clean, dev-other)	CC BY 4.0	Panayotov et al. (2015); audio from LibriVox	https://www.openslr.org/12/
FLEURS (en_us)	CC BY 4.0	Conneau et al. (2023)	https://huggingface.co/datasets/google/fleurs (card metadata license: cc-by-4.0)
Common Voice 9.0 English (set excluded from this repo)	CC0 1.0, with Mozilla's Common Voice terms	Ardila et al. (2020)	https://github.com/mozilla/legal-docs (prod/en/common_voice_terms.md)
VoxPopuli (English, ESB copy)	CC0 1.0 (data); European Parliament legal notice for the raw data	Wang et al. (2021); source of the raw recordings: European Parliament	https://github.com/facebookresearch/voxpopuli and https://www.europarl.europa.eu/legal-notice/en/
AMI Meeting Corpus (IHM, ESB copy)	CC BY 4.0	Carletta et al. (2006)	https://groups.inf.ed.ac.uk/ami/corpus/license.shtml
Multilingual LibriSpeech (English)	CC BY 4.0	Pratap et al. (2020)	https://www.openslr.org/94/

• LibriSpeech word and phone alignments (alignments/librispeech-dev_*.parquet) are derived from the LibriSpeech Alignments of Lugosch (Zenodo 10.5281/zenodo.2619474), licensed CC BY 4.0 (https://zenodo.org/records/2619474). The record asks users to cite Lugosch et al. (2019) and the Montreal Forced Aligner paper (McAuliffe et al., 2017).
• The European Parliament legal notice allows reuse, including commercial reuse, provided the source is acknowledged, and asks that partial reproductions cite the URL of the complete item. The recordings come from European Parliament plenary sessions (https://www.europarl.europa.eu/plenary/en/debates-video.html).

The scripts, reports and build metadata written for this repo are released under CC BY 4.0. The models used to compute the units have their own licenses and this repo contains none of their weights: HuBERT Large (Apache 2.0), SpireKMeans (CC BY-NC 4.0), SpeechTokenizer (Apache 2.0), Mimi (CC BY 4.0) and FocalCodec (Apache 2.0). The HuBERT units were assigned with the non-commercial SpireKMeans centroids. Whether that license term extends to the unit indices is not settled by the model license, so check it before commercial use.

Citation

LibriSpeech (Panayotov et al., ICASSP 2015):

@inproceedings{panayotov2015librispeech,
  title     = {Librispeech: An {ASR} corpus based on public domain audio books},
  author    = {Panayotov, Vassil and Chen, Guoguo and Povey, Daniel and Khudanpur, Sanjeev},
  booktitle = {2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages     = {5206--5210},
  year      = {2015},
  publisher = {IEEE},
  doi       = {10.1109/ICASSP.2015.7178964}
}

FLEURS (Conneau et al., IEEE SLT 2022):

@inproceedings{conneau2023fleurs,
  title     = {{FLEURS}: Few-Shot Learning Evaluation of Universal Representations of Speech},
  author    = {Conneau, Alexis and Ma, Min and Khanuja, Simran and Zhang, Yu and Axelrod, Vera and Dalmia, Siddharth and Riesa, Jason and Rivera, Clara and Bapna, Ankur},
  booktitle = {2022 IEEE Spoken Language Technology Workshop (SLT)},
  pages     = {798--805},
  year      = {2023},
  publisher = {IEEE},
  doi       = {10.1109/SLT54892.2023.10023141},
  eprint    = {2205.12446},
  archivePrefix = {arXiv}
}

VoxPopuli (Wang et al., ACL 2021):

@inproceedings{wang2021voxpopuli,
  title     = {{V}ox{P}opuli: A Large-Scale Multilingual Speech Corpus for Representation Learning, Semi-Supervised Learning and Interpretation},
  author    = {Wang, Changhan and Riviere, Morgane and Lee, Ann and Wu, Anne and Talnikar, Chaitanya and Haziza, Daniel and Williamson, Mary and Pino, Juan and Dupoux, Emmanuel},
  booktitle = {Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers)},
  pages     = {993--1003},
  month     = aug,
  year      = {2021},
  address   = {Online},
  publisher = {Association for Computational Linguistics},
  doi       = {10.18653/v1/2021.acl-long.80},
  eprint    = {2101.00390},
  archivePrefix = {arXiv}
}

AMI Meeting Corpus (Carletta et al., MLMI 2005, LNCS 3869, 2006):

@inproceedings{carletta2006ami,
  title     = {The {AMI} Meeting Corpus: A Pre-announcement},
  author    = {Carletta, Jean and Ashby, Simone and Bourban, Sebastien and Flynn, Mike and Guillemot, Mael and Hain, Thomas and Kadlec, Jaroslav and Karaiskos, Vasilis and Kraaij, Wessel and Kronenthal, Melissa and Lathoud, Guillaume and Lincoln, Mike and Lisowska, Agnes and McCowan, Iain and Post, Wilfried and Reidsma, Dennis and Wellner, Pierre},
  editor    = {Renals, Steve and Bengio, Samy},
  booktitle = {Machine Learning for Multimodal Interaction (MLMI 2005)},
  series    = {Lecture Notes in Computer Science},
  volume    = {3869},
  pages     = {28--39},
  year      = {2006},
  publisher = {Springer Berlin Heidelberg},
  doi       = {10.1007/11677482_3}
}

ESB, the source of the Open ASR Leaderboard copies (Gandhi et al., arXiv 2022):

@misc{gandhi2022esb,
  title  = {{ESB}: A Benchmark For Multi-Domain End-to-End Speech Recognition},
  author = {Gandhi, Sanchit and von Platen, Patrick and Rush, Alexander M.},
  year   = {2022},
  eprint = {2210.13352},
  archivePrefix = {arXiv},
  doi    = {10.48550/arXiv.2210.13352}
}

MLS (Pratap et al., Interspeech 2020):

@inproceedings{pratap2020mls,
  title     = {{MLS}: A Large-Scale Multilingual Dataset for Speech Research},
  author    = {Pratap, Vineel and Xu, Qiantong and Sriram, Anuroop and Synnaeve, Gabriel and Collobert, Ronan},
  booktitle = {Interspeech 2020},
  pages     = {2757--2761},
  year      = {2020},
  publisher = {ISCA},
  doi       = {10.21437/Interspeech.2020-2826},
  eprint    = {2012.03411},
  archivePrefix = {arXiv}
}

LibriSpeech alignments (the Zenodo record asks for both the paper and the record) (Lugosch et al., Interspeech 2019; Zenodo 10.5281/zenodo.2619474):

@inproceedings{lugosch2019slu,
  title     = {Speech Model Pre-Training for End-to-End Spoken Language Understanding},
  author    = {Lugosch, Loren and Ravanelli, Mirco and Ignoto, Patrick and Tomar, Vikrant Singh and Bengio, Yoshua},
  booktitle = {Interspeech 2019},
  pages     = {814--818},
  year      = {2019},
  publisher = {ISCA},
  doi       = {10.21437/Interspeech.2019-2396},
  eprint    = {1904.03670},
  archivePrefix = {arXiv}
}

@misc{lugosch2019librispeechalignments,
  title     = {{LibriSpeech} Alignments},
  author    = {Lugosch, Loren},
  year      = {2019},
  version   = {1.0},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.2619474},
  url       = {https://zenodo.org/records/2619474}
}

Montreal Forced Aligner (McAuliffe et al., Interspeech 2017):

@inproceedings{mcauliffe2017mfa,
  title     = {Montreal Forced Aligner: Trainable Text-Speech Alignment Using {Kaldi}},
  author    = {McAuliffe, Michael and Socolof, Michaela and Mihuc, Sarah and Wagner, Michael and Sonderegger, Morgan},
  booktitle = {Interspeech 2017},
  pages     = {498--502},
  year      = {2017},
  publisher = {ISCA},
  doi       = {10.21437/Interspeech.2017-1386}
}

HuBERT (Hsu et al., IEEE/ACM TASLP 2021):

@article{hsu2021hubert,
  title   = {{HuBERT}: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units},
  author  = {Hsu, Wei-Ning and Bolte, Benjamin and Tsai, Yao-Hung Hubert and Lakhotia, Kushal and Salakhutdinov, Ruslan and Mohamed, Abdelrahman},
  journal = {IEEE/ACM Transactions on Audio, Speech, and Language Processing},
  volume  = {29},
  pages   = {3451--3460},
  year    = {2021},
  doi     = {10.1109/TASLP.2021.3122291},
  eprint  = {2106.07447},
  archivePrefix = {arXiv}
}

SPIRE, the source of the SpireKMeans k-means model (Ambilduke et al., Findings of EMNLP 2025):

@inproceedings{ambilduke2025spire,
  title     = {From Tower to Spire: Adding the Speech Modality to a Translation-Specialist {LLM}},
  author    = {Ambilduke, Kshitij and Peters, Ben and Sannigrahi, Sonal and Keshwani, Anil and Lam, Tsz Kin and Martins, Bruno and Martins, Andre and Zanon Boito, Marcely},
  editor    = {Christodoulopoulos, Christos and Chakraborty, Tanmoy and Rose, Carolyn and Peng, Violet},
  booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2025},
  pages     = {19658--19673},
  month     = nov,
  year      = {2025},
  address   = {Suzhou, China},
  publisher = {Association for Computational Linguistics},
  doi       = {10.18653/v1/2025.findings-emnlp.1071},
  url       = {https://aclanthology.org/2025.findings-emnlp.1071/},
  eprint    = {2503.10620},
  archivePrefix = {arXiv}
}

SpeechTokenizer (Zhang et al., ICLR 2024):

@inproceedings{zhang2024speechtokenizer,
  title     = {{SpeechTokenizer}: Unified Speech Tokenizer for Speech Language Models},
  author    = {Zhang, Xin and Zhang, Dong and Li, Shimin and Zhou, Yaqian and Qiu, Xipeng},
  booktitle = {The Twelfth International Conference on Learning Representations (ICLR)},
  year      = {2024},
  url       = {https://openreview.net/forum?id=AF9Q8Vip84},
  eprint    = {2308.16692},
  archivePrefix = {arXiv}
}

Mimi (introduced with Moshi) (Défossez et al., arXiv 2024):

@misc{defossez2024moshi,
  title  = {Moshi: a speech-text foundation model for real-time dialogue},
  author = {D{\'e}fossez, Alexandre and Mazar{\'e}, Laurent and Orsini, Manu and Royer, Am{\'e}lie and P{\'e}rez, Patrick and J{\'e}gou, Herv{\'e} and Grave, Edouard and Zeghidour, Neil},
  year   = {2024},
  eprint = {2410.00037},
  archivePrefix = {arXiv},
  primaryClass  = {eess.AS},
  doi    = {10.48550/arXiv.2410.00037}
}

FocalCodec (Della Libera et al., NeurIPS 2025):

@inproceedings{dellalibera2025focalcodec,
  title     = {{FocalCodec}: Low-Bitrate Speech Coding via Focal Modulation Networks},
  author    = {Della Libera, Luca and Paissan, Francesco and Subakan, Cem and Ravanelli, Mirco},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2025},
  url       = {https://openreview.net/forum?id=7Z3wQSu3mH},
  eprint    = {2502.04465},
  archivePrefix = {arXiv}
}

Related repos

• anilkeshwani/speech-integration-speech-bpe
• anilkeshwani/mls-hubert_large_ll60k-layer_22
• anilkeshwani/mls-speechtokenizer-rvq_0
• anilkeshwani/mls-mimi-srvq_0
• anilkeshwani/mls-focalcodec
• Collection: Speech units for LLM speech integration (MLS English)