Elamur Research⌕
☰
home / papers / 2026 / Supplementary Resources and Analysis for Automatic Speech Recognition Systems Trained on the Loquacious Dataset

Supplementary Resources and Analysis for Automatic Speech Recognition Systems Trained on the Loquacious Dataset

Summary

The Loquacious dataset is a large, openly licensed English ASR corpus intended to replace older benchmarks such as LibriSpeech and TED-Lium. Its original release contained only audio, transcripts and train/dev/test splits. It lacked the auxiliary material that many classical and phoneme-based pipelines need. The authors from RWTH Aachen and AppTek fill this gap. They release a 216k-word vocabulary, a phoneme pronunciation lexicon derived from CMUdict, a Sequitur G2P model for out-of-dictionary words, and several KenLM n-gram language models. Using these resources, they train Conformer-based systems of several types: CTC, RNN-T, a phoneme-based factored hybrid (FH) and AED. Training uses the 250-hour and 2.5k-hour subsets. They compare BPE and phoneme units, greedy and beam search, lexicon-constrained decoding with and without a 4-gram LM, pronunciation-variant settings, and data augmentation. The study gives the community reproducible reference baselines for Loquacious. It also shows that the dataset exposes typical ASR challenges, such as domain differences between its test sources.

Key points

  • Released resources: All resources are on HuggingFace (speechbrain/LoquaciousSet), and the generation scripts are on GitHub (rwth-i6/LoquaciousAdditionalResources).

  • Vocabulary construction: The 216k-word vocabulary combines two sets:

    • 95k CMUdict 0.7b words that appear in the train-large text;
    • 197k words that occur at least 4 times in train-large.

  • Lexicon and G2P: CMUdict pronunciations are used with stress removed. Words missing from CMUdict get pronunciations from a 5th-order Sequitur G2P model. A second variant is added only when the top variant's posterior falls below a 60% or 80% threshold.

  • Language models: KenLM builds three ARPA models:

    • a pruned 3-gram (36M n-grams);
    • a pruned 4-gram (54M n-grams);
    • an unpruned 4-gram (246M n-grams).

    Dev-set OOV rates stay roughly between 0.1% and 1.3%, and perplexity differs noticeably by source (CommonVoice, LibriSpeech, VoxPopuli, Yodas).

  • Experimental setup: All systems share a 12-layer, 512-dimensional Conformer encoder. The CTC model has about 77M parameters and the RNN-T about 80M. Models train for 100 epochs on train.small and 40 epochs on train.medium, with SpecAugment and speed perturbation (the FH model uses SpecAugment only).

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Linguistics & languages

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/4zsvhm25r7zf

@inproceedings{rossenbach2026supplementary,
  title = {{Supplementary Resources and Analysis for Automatic Speech Recognition Systems Trained on the Loquacious Dataset}},
  author = {Rossenbach, Nick and Schmitt, Robin and Raissi, Tina and Berger, Simon and Kleppel, Larissa and Schlüter, Ralf},
  booktitle = {Proceedings of the Fifteenth Language Resources and Evaluation Conference},
  year = {2026},
  doi = {10.63317/4zsvhm25r7zf},
  url = {https://doi.org/10.63317/4zsvhm25r7zf}
}

References

29 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Vladimir Bataev, Andrei Andrusenko, Lilit Grigoryan, Aleksandr Laptev, Vitaly Lavrukhin, and Boris Ginsburg. 2025. NGPU-LM: GPU-Accelerated N-Gram Language Model for Context-Biasing in Greedy ASR Decoding. In Proc. Interspeech, pages 644–648. find
  2. Maximilian Bisani and Hermann Ney. 2008. Joint-sequence models for grapheme-to-phoneme conversion. Speech Communication, 50(5):434– 451. find
  3. Hervé A. Bourlard and Nelson Morgan. 1993. Connectionist Speech Recognition: a Hybrid Approach. Kluwer Academic Publishers, Norwell, MA. find
  4. William Chan, Navdeep Jaitly, Quoc Le, and Oriol Vinyals. 2016. Listen, attend and spell: A neural network for large vocabulary conversational speech recognition. In Proc. IEEE ICASSP, pages 4960–4964. find
  5. Rita Frieske and Bertram E. Shi. 2024. Hallucinations in neural automatic speech recognition: Identifying errors and hallucinatory models. Preprint arXiv: 2401.01572. — [arXiv:2401.01572](https://arxiv.org/abs/2401.01572) find
  6. J.J. Godfrey, E.C. Holliman, and J. McDaniel. 1992. Switchboard: telephone speech corpus for research and development. In Proc. IEEE ICASSP, volume 1, pages 517–520 vol.1. find
  7. Alex Graves. 2012. Sequence transduction with recurrent neural networks. In RL 2012 - ICML Workshop on Representation Learning. find our card
  8. Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhuber. 2006. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. ICML ’06, page 369–376, New York, NY, USA. find
  9. Anmol Gulati, James Qin, Chung-Cheng Chiu, Niki Parmar, Yu Zhang, Jiahui Yu, Wei Han, Shibo Wang, Zhengdong Zhang, Yonghui Wu, and Ruoming Pang. 2020. Conformer: Convolution-augmented transformer for speech recognition. In Proc. Interspeech, pages 5036–5040. find our card
  10. Kenneth Heafield. 2011. KenLM: Faster and smaller language model queries. In Proceedings of the Sixth Workshop on Statistical Machine Translation, pages 187–197, Edinburgh, Scotland. Proc. ACL. find our card
  11. François Hernandez, Vincent Nguyen, Sahar Ghannay, Natalia A. Tomashenko, and Yannick Estève. 2018. TED-LIUM 3: Twice as much data and corpus repartition for experiments on speaker adaptation. In Speech and Computer - 20th International Conference, SPECOM 2018, Leipzig, Germany, September 18-22, 201… find
  12. Dongcheng Jiang, Chao Zhang, and Philip C Woodland. 2023. A neural time alignment module for end-to-end automatic speech recognition. In Proc. Interspeech, pages 1374–1378. find
  13. Jacob Kahn, Vineel Pratap, Tatiana Likhomanenko, Qiantong Xu, Awni Hannun, Jeff Cai, Paden Tomasello, Ann Lee, Edouard Grave, Gilad Avidov, Benoit Steiner, Vitaliy Liptchinsky, Gabriel Synnaeve, and Ronan Collobert. 2022. Flashlight: Enabling innovation in tools for machine learning. Preprint arXiv:… — [arXiv:2201.12465](https://arxiv.org/abs/2201.12465) find
  14. Tom Ko, Vijayaditya Peddinti, Daniel Povey, and Sanjeev Khudanpur. 2015. Audio augmentation for speech recognition. In Proc. Interspeech, pages 3586–3589. find
  15. Taku Kudo. 2018. Subword regularization: Improving neural network translation models with multiple subword candidates. In Proc. ACL, pages 66–75, Melbourne, Australia. find
  16. Zhong Meng, Sarangarajan Parthasarathy, Eric Sun, Yashesh Gaur, Naoyuki Kanda, Liang Lu, Xie Chen, Rui Zhao, Jinyu Li, and Yifan Gong. 2021. Internal language model estimation for domain-adaptive end-to-end speech recognition. In Proc. IEEE SLT, pages 243–250. find
  17. Mehryar Mohri, Fernando Pereira, and Michael Riley. 2002. Weighted finite-state transducers in speech recognition. Computer Speech & Language, 16(1):69–88. find
  18. Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. 2015. Librispeech: An asr corpus based on public domain audio books. In Proc. IEEE ICASSP, pages 5206–5210. find
  19. Titouan Parcollet, Yuan Tseng, Shucong Zhang, and Rogier C. van Dalen. 2025. Loquacious Set: 25,000 Hours of Transcribed and Diverse English Speech Recognition Data for Research and Commercial Use. In Proc. Interspeech, pages 4053–4057. find
  20. Daniel S. Park, William Chan, Yu Zhang, Chung-Cheng Chiu, Barret Zoph, Ekin D. Cubuk, and Quoc V. Le. 2019. Specaugment: A simple data augmentation method for automatic speech recognition. Preprint arXiv: 1904.08779. find our card
  21. Jan-Thorsten Peter, Eugen Beck, and Hermann Ney. 2018. Sisyphus, a workflow manager designed for machine translation and automatic speech recognition. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, EMNLP, pages 84–89. find
  22. Tina Raissi, Eugen Beck, Ralf Schlüter, and Hermann Ney. 2020. Context-dependent acoustic modeling without explicit phone clustering. In Proc. Interspeech, pages 4377–4381. find
  23. Tina Raissi, Nick Rossenbach, and Ralf Schlüter. 2025. Analysis of Domain Shift across ASR Architectures via TTS-Enabled Separation of Target Domain and Acoustic Conditions. In Proc. IEEE ASRU. find
  24. Rotem Rousso, Eyal Cohen, Joseph Keshet, and Eleanor Chodroff. 2024. Tradition or Innovation: A Comparison of Modern ASR Methods for Forced Alignment. In Proc. Interspeech, pages 1525–1529. find
  25. David Rybach, Stefan Hahn, Patrick Lehnen, David Nolden, Martin Sundermeyer, Zoltan Tüske, Simon Wiesler, Ralf Schlüter, and Hermann Ney. 2011. RASR-the RWTH Aachen university open source speech recognition toolkit. In Proc. IEEE ASRU. find
  26. Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Neural machine translation of rare words with subword units. In Proc. ACL, pages 1715– 1725. find
  27. Anshuman Tripathi, Han Lu, Hasim Sak, and Hagen Soltau. 2019. Monotonic recurrent neural network transducer and decoding strategies. In Proc. IEEE ASRU, pages 944–948. find
  28. Ehsan Variani, David Rybach, Cyril Allauzen, and Michael Riley. 2020. Hybrid autoregressive transducer (hat). In Proc. IEEE ICASSP, pages 6139– 6143. find
  29. Albert Zeyer, Tamer Alkhouli, and Hermann Ney. 2018. RETURNN as a Generic Flexible Neural Toolkit with Application to Translation and Speech Recognition. In Proc. ACL, pages 128– 133, Melbourne, Australia. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/rossenbach-2026-supplementary-resources-analysis-automatic-speech/