---
title: "English Broadcast News Speech Recognition by Humans and"
description: "The authors therefore build reliable training subsets: they decode the captioned audio with several ASR systems, bias each… 21 references."
authors: ["Samuel Thomas", "Masayuki Suzuki", "Yinghui Huang", "Gakuto Kurata", "Zoltan Tuske", "George Saon", "Brian Kingsbury", "Michael Picheny", "Tom Dibert", "Alice Kaiser-Schatzlein", "Bern Samko"]
year: 2019
venue: "arXiv:1904.13258"
source: "https://arxiv.org/abs/1904.13258"
doi: "10.1109/icassp.2019.8683211"
arxiv: "1904.13258"
licence: "arXiv — non-exclusive license to distribute"
licence_url: "http://arxiv.org/licenses/nonexclusive-distrib/1.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/thomas-2019-english-broadcast-news-speech-recognition/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/thomas-2019-english-broadcast-news-speech-recognition/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2019](https://md.research.elamur.ai/free/en/papers/2019/page/6/) / English Broadcast News Speech Recognition by Humans and Machines

# English Broadcast News Speech Recognition by Humans and Machines

**Authors:** [Samuel Thomas](https://md.research.elamur.ai/free/en/authors/samuel-thomas/), Masayuki Suzuki, Yinghui Huang, [Gakuto Kurata](https://md.research.elamur.ai/free/en/authors/gakuto-kurata/), [Zoltan Tuske](https://md.research.elamur.ai/free/en/authors/zoltan-tuske/), [George Saon](https://md.research.elamur.ai/free/en/authors/george-saon/), [Brian Kingsbury](https://md.research.elamur.ai/free/en/authors/brian-kingsbury/), [Michael Picheny](https://md.research.elamur.ai/free/en/authors/michael-picheny/), Tom Dibert, Alice Kaiser-Schatzlein and Bern Samko · **Year:** 2019 · **Venue:** [arXiv:1904.13258](https://md.research.elamur.ai/free/en/venues/ieee-international-conference-on-acoustics-speech-and-signal-processing/) ([IEEE International Conference on Acoustics Speech and Signal Processing 2019](https://md.research.elamur.ai/free/en/venues/ieee-international-conference-on-acoustics-speech-and-signal-processing/2019/)) · **Licence:** [arXiv — non-exclusive license to distribute](http://arxiv.org/licenses/nonexclusive-distrib/1.0/)

**Links:** [Publisher's page](https://doi.org/10.1109/icassp.2019.8683211) · [PDF](https://arxiv.org/pdf/1904.13258) · [DOI 10.1109/icassp.2019.8683211](https://doi.org/10.1109/icassp.2019.8683211) · [arXiv 1904.13258](https://arxiv.org/abs/1904.13258) · [OpenAlex](https://openalex.org/W2937657912)

**Citations:** 12 citations (OpenAlex, 2026-10-14)

## Summary
The paper asks whether deep-learning recipes that brought conversational telephone speech (CTS) recognition near human level also work for English broadcast news (BN). It also measures how far machines still are from people on this task. BN has little carefully transcribed audio, about 140 hours. Most of the roughly 9000 hours available come only with closed captions. The authors therefore build reliable training subsets: they decode the captioned audio with several ASR systems, bias each decode toward the captions, and keep only material where confidence is high and the systems agree. They train LSTM and ResNet acoustic models and decode with n-gram plus feed-forward neural LMs, followed by LSTM LM rescoring. The final word error rates are 6.5% on DEV04F and 5.9% on RT04, new results for these sets. Professional transcribers hired through Appen reach 3.6% and 2.8% under the same scoring setup. The CTS techniques transfer well to BN, but a clear gap to human performance remains.

## Key points
- Test sets are DEV04F (about 2 h, 6 shows) and RT04 (about 4 h, 12 shows). Scoring uses NIST sclite with the original EARS references and scripts.
- Training corpora:
  - BN-400 (~430 h) combines the 144 h of careful Hub4 transcripts with filtered semi-supervised data.
  - BN-1300 adds about 900 h more to BN-400.
  - The sources are TDT2/TDT4 and BN03 closed-captioned audio.
- The LSTM acoustic model has 6 bidirectional layers with 1024 cells each, a 256-unit bottleneck and 32K context-dependent HMM state outputs. It is trained with cross-entropy on BN-1300, then sequence-trained on the 144 h clean set only.
- WER progression on DEV04F / RT04:
  - LSTM with n-gram LM: 7.6 / 7.7.
  - Adding the FFNN-LM: 7.2 / 7.0.
  - Combined LSTM1/LSTM2 rescoring: 6.5 / 5.9.
  - ResNet alone is weaker (9.0 / 8.1 with FFNN-LM).
- Human protocol:
  - Three independent transcribers took 14–16×RT each, and a fourth senior transcriber checked quality at about 3×RT.
  - Individual human WERs ranged from 3.6% to 4.4% on DEV04F and from 2.8% to 3.6% on RT04.
  - Transcribers found BN dense, noisy and full of named entities that needed lookup.

## Interesting ideas
- Noisy captions become usable labels by decoding with several independent recognizers, each nudged toward the caption, and keeping only segments where they agree confidently; agreement between systems acts as the quality filter.
- Training runs in two phases: a large, filtered but imperfect corpus does the bulk learning, and only the small gold set is used for a final discriminative refinement.
- The human baseline is measured, not assumed: several annotators follow the original guidelines, a reviewer checks their work, and the output goes through the same normalization and scorer as the machine. This gives a realistic ceiling and shows the spread between people.
- Two architecturally different models with comparable accuracy are kept together because their errors differ, even when one model alone is weaker.

## For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

- **I1** (learn): The same rule fits silver treebank building: keep only sentences where the witness parsers agree on arcs and labels with high confidence, and send the rest to an LLM teacher tier.
- **I2** (learn): Train the perceptron parser or gparse on silver data first, then do a last fine-tuning pass restricted to hand-checked UD treebank sentences, so gold data shapes the final decision boundaries.
- **I3** (answer): Have several people answer the school tests and fable QA under the same scoring script, so Mova's accuracy is judged against measured human variance.
- **I4** (tree): Keep both the perceptron parser and gparse and study where their trees disagree. Disagreement can flag hard constructions and support explainable arc-level voting.

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

## Institutions

- [IBM](https://md.research.elamur.ai/free/en/companies/ibm/) — Company, Armonk, United States · 85 works · 66,118 citations

Also on the byline: [IBM Research - Tokyo](https://ror.org/04915qk43).

📚 Shelves (5): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 9,086 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 8,889 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 9,296 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 9,206 works · [🔊 Signal Processing](https://md.research.elamur.ai/free/en/shelves/computer-science/signal-processing/) · 213 works

🧭 Topics (2): [English](https://md.research.elamur.ai/free/en/topics/english/) · 1,373 works · [🎙️ Speech & spoken language](https://md.research.elamur.ai/free/en/topics/speech/) · 233 works

## Cite

Cite as: Samuel Thomas, Masayuki Suzuki, Yinghui Huang, Gakuto Kurata, Zoltan Tuske, George Saon, Brian Kingsbury, Michael Picheny, Tom Dibert, Alice Kaiser-Schatzlein and Bern Samko. 2019. English Broadcast News Speech Recognition by Humans and Machines. arXiv:1904.13258.

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@misc{thomas2019english,
  title = {{English Broadcast News Speech Recognition by Humans and Machines}},
  author = {Samuel Thomas, Masayuki Suzuki, Yinghui Huang, Gakuto Kurata, Zoltan Tuske, George Saon, Brian Kingsbury, Michael Picheny, Tom Dibert, Alice Kaiser-Schatzlein and Bern Samko},
  year = {2019},
  doi = {10.1109/icassp.2019.8683211},
  eprint = {1904.13258},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.1109/icassp.2019.8683211}
}
```

- [← previous Amharic-Arabic Neural Machine Translation](https://md.research.elamur.ai/free/en/papers/gashaw-2019-amharic-arabic-neural-machine-translation/)
- [next → Restoring Arabic vowels through omission-tolerant dictionary lookup](https://md.research.elamur.ai/free/en/papers/neme-2019-restoring-arabic-vowels-through-omission/)
- [↑ back to papers of 2019](https://md.research.elamur.ai/free/en/papers/2019/page/6/)

## References

21 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Joseph Olive, Caitlin Christianson, and John McCary, Handbook of natural language processing and machine translation: DARPA Global Autonomous Language Exploitation, Springer Science & Business Media, 2011. [find](https://scholar.google.com/scholar?q=Joseph+Olive%2C+Caitlin+Christianson%2C+and+John+McCary%2C+Handbook+of+natural+language+processing+and+machine+translation%3A+DARPA+Global+Autonomous+Language+Exploitation%2C+Springer+Science+%26+Business+Media%2C+)
2. George Saon, Gakuto Kurata, Tom Sercu, Kartik Audhkhasi, Samuel Thomas, Dimitrios Dimitriadis, Xiaodong Cui, Bhuvana Ramabhadran, Michael Picheny, Lynn-Li Lim, et al., “[English conversational telephone speech recognition by humans and machines](https://md.research.elamur.ai/free/en/papers/saon-2017-english-conversational-telephone-speech-recognition/),” arXiv preprint arXiv:1703.02136, 2017. [↗](https://arxiv.org/abs/1703.02136)
3. Gakuto Kurata, Bhuvana Ramabhadran, George Saon, and Abhinav Sethy, “Language modeling with highway LSTM,” in Automatic Speech Recognition and Understanding Workshop (ASRU), 2017 IEEE. IEEE, 2017, pp. 244–251. [find](https://scholar.google.com/scholar?q=Gakuto+Kurata%2C+Bhuvana+Ramabhadran%2C+George+Saon%2C+and+Abhinav+Sethy%2C+%E2%80%9CLanguage+modeling+with+highway+LSTM%2C%E2%80%9D+in+Automatic+Speech+Recognition+and+Understanding+Workshop+%28ASRU%29%2C+2017+IEEE%2E+IEEE%2C+2017%2C+pp%2E)
4. Wayne Xiong, Lingfeng Wu, Fil Alleva, Jasha Droppo, Xuedong Huang, and Andreas Stolcke, “The Microsoft 2017 conversational speech recognition system,” in 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2018, pp. 5934–5938. [find](https://scholar.google.com/scholar?q=Wayne+Xiong%2C+Lingfeng+Wu%2C+Fil+Alleva%2C+Jasha+Droppo%2C+Xuedong+Huang%2C+and+Andreas+Stolcke%2C+%E2%80%9CThe+Microsoft+2017+conversational+speech+recognition+system%2C%E2%80%9D+in+2018+IEEE+International+Conference+on+Acoustic)
5. Kyu J Han, Akshay Chandrashekaran, Jungsuk Kim, and Ian Lane, “The CAPIO 2017 conversational speech recognition system,” arXiv preprint arXiv:1801.00059, 2017. [↗](https://arxiv.org/abs/1801.00059)
6. Lori Lamel, Jean-Luc Gauvain, and Gilles Adda, “Lightly supervised acoustic model training,” in ASR2000-Automatic Speech Recognition: Challenges for the new Millenium ISCA Tutorial and Research Workshop (ITRW), 2000. [find](https://scholar.google.com/scholar?q=Lori+Lamel%2C+Jean%2DLuc+Gauvain%2C+and+Gilles+Adda%2C+%E2%80%9CLightly+supervised+acoustic+model+training%2C%E2%80%9D+in+ASR2000%2DAutomatic+Speech+Recognition%3A+Challenges+for+the+new+Millenium+ISCA+Tutorial+and+Research+Worksh)
7. Mark JF Gales, PC Woodland, Ho Yin Chan, David Mrva, Rohit Sinha, Sue E Tranter, et al., “Progress in the CU-HTK broadcast news transcription system,” IEEE Transactions on Audio, Speech, and Language Processing, vol. 14, no. 5, pp. 1513–1525, 2006. [↗](https://doi.org/10.1109/tasl.2006.878264)
8. Spyridon Matsoukas, J-L Gauvain, Gilles Adda, Thomas Colthurst, Chia-Lin Kao, Owen Kimball, Lori Lamel, Fabrice Lefevre, Jeff Z Ma, John Makhoul, et al., “Advances in transcription of broadcast news and conversational telephone speech within the combined EARS BBN/LIMSI system,” IEEE Transactions on Audio, Speech, and Language Processing, vol. 14, no. 5, pp. 1541–1556, 2006. [↗](https://doi.org/10.1109/tasl.2006.878257)
9. Stanley F Chen, Brian Kingsbury, Lidia Mangu, Daniel Povey, George Saon, Hagen Soltau, and Geoffrey Zweig, “Advances in speech transcription at IBM under the DARPA EARS program,” IEEE Transactions on Audio, Speech, and Language Processing, vol. 14, no. 5, pp. 1596–1608, 2006. [↗](https://doi.org/10.1109/tasl.2006.879814)
10. Andreas Stolcke, Barry Chen, Horacio Franco, Venkata Ramana Rao Gadde, Martin Graciarena, Mei-Yuh Hwang, Katrin Kirchhoff, Arindam Mandal, Nelson Morgan, Xin Lei, et al., “Recent innovations in speech-to-text transcription at SRI-ICSI-UW,” IEEE Transactions on Audio, Speech, and Language Processing, vol. 14, no. 5, pp. 1729–1744, 2006. [↗](https://doi.org/10.1109/tasl.2006.879807)
11. Rohit Sinha, Mark JF Gales, DY Kim, X Andrew Liu, Khe Chai Sim, and Philip C Woodland, “The CU-HTK Mandarin broadcast news transcription system,” in Acoustics, Speech and Signal Processing, 2006. ICASSP 2006 Proceedings. 2006 IEEE International Conference on. IEEE, 2006, vol. 1, pp. I–I. [find](https://scholar.google.com/scholar?q=Rohit+Sinha%2C+Mark+JF+Gales%2C+DY+Kim%2C+X+Andrew+Liu%2C+Khe+Chai+Sim%2C+and+Philip+C+Woodland%2C+%E2%80%9CThe+CU%2DHTK+Mandarin+broadcast+news+transcription+system%2C%E2%80%9D+in+Acoustics%2C+Speech+and+Signal+Processing%2C+2006%2E+ICAS)
12. Xin Lei, Manhung Siu, Mei-Yuh Hwang, Mari Ostendorf, and Tan Lee, “Improved tone modeling for Mandarin broadcast news speech recognition,” in Ninth International Conference on Spoken Language Processing, 2006. [find](https://scholar.google.com/scholar?q=Xin+Lei%2C+Manhung+Siu%2C+Mei%2DYuh+Hwang%2C+Mari+Ostendorf%2C+and+Tan+Lee%2C+%E2%80%9CImproved+tone+modeling+for+Mandarin+broadcast+news+speech+recognition%2C%E2%80%9D+in+Ninth+International+Conference+on+Spoken+Language+Processi)
13. Bing Xiang, Kham Nguyen, Long Nguyen, Richard Schwartz, and John Makhoul, “Morphological decomposition for Arabic broadcast news transcription,” in Acoustics, Speech and Signal Processing, 2006. ICASSP 2006 Proceedings. 2006 IEEE International Conference on. IEEE, 2006, vol. 1, pp. I–I. [find](https://scholar.google.com/scholar?q=Bing+Xiang%2C+Kham+Nguyen%2C+Long+Nguyen%2C+Richard+Schwartz%2C+and+John+Makhoul%2C+%E2%80%9CMorphological+decomposition+for+Arabic+broadcast+news+transcription%2C%E2%80%9D+in+Acoustics%2C+Speech+and+Signal+Processing%2C+2006%2E+ICASS)
14. Peter Bell, Mark JF Gales, Thomas Hain, Jonathan Kilgour, Pierre Lanchantin, Xunying Liu, Andrew McParland, Steve Renals, Oscar Saz, Mirjam Wester, et al., “The MGB challenge: Evaluating multi-genre broadcast media recognition,” in Automatic Speech Recognition and Understanding (ASRU), 2015 IEEE Workshop on. IEEE, 2015, pp. 687–693. [find](https://scholar.google.com/scholar?q=Peter+Bell%2C+Mark+JF+Gales%2C+Thomas+Hain%2C+Jonathan+Kilgour%2C+Pierre+Lanchantin%2C+Xunying+Liu%2C+Andrew+McParland%2C+Steve+Renals%2C+Oscar+Saz%2C+Mirjam+Wester%2C+et+al%2E%2C+%E2%80%9CThe+MGB+challenge%3A+Evaluating+multi%2Dgenre+b)
15. Ahmed Ali, Stephan Vogel, and Steve Renals, “[Speech recognition challenge in the wild: Arabic MGB-3](https://md.research.elamur.ai/free/en/papers/ali-2017-speech-recognition-challenge-wild-arabic/),” in Automatic Speech Recognition and Understanding Workshop (ASRU), 2017 IEEE. IEEE, 2017, pp. 316–322.
16. Yinghui Huang, Abhinav Sethy, and Bhuvana Ramabhadran, “Fast neural network language model lookups at n-gram speeds,” in Interspeech 2017. ISCA, 2017. [find](https://scholar.google.com/scholar?q=Yinghui+Huang%2C+Abhinav+Sethy%2C+and+Bhuvana+Ramabhadran%2C+%E2%80%9CFast+neural+network+language+model+lookups+at+n%2Dgram+speeds%2C%E2%80%9D+in+Interspeech+2017%2E+ISCA%2C+2017%2E)
17. Abhinav Sethy, Stanley Chen, Ebru Arisoy, and Bhuvana Ramabhadran, “Unnormalized exponential and neural network language models,” in ICASSP, 2015. [find](https://scholar.google.com/scholar?q=Abhinav+Sethy%2C+Stanley+Chen%2C+Ebru+Arisoy%2C+and+Bhuvana+Ramabhadran%2C+%E2%80%9CUnnormalized+exponential+and+neural+network+language+models%2C%E2%80%9D+in+ICASSP%2C+2015%2E)
18. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778. [↗](https://doi.org/10.48550/arxiv.1512.03385)
19. Diederik P Kingma and Jimmy Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014. [↗](https://arxiv.org/abs/1412.6980)
20. Stephen Merity, Nitish Shirish Keskar, and Richard Socher, “[Regularizing and optimizing LSTM language models](https://md.research.elamur.ai/free/en/papers/merity-2017-regularizing-optimizing-lstm-language-models/),” in International Conference on Learning Representations (ICLR), 2018. [↗](https://doi.org/10.48550/arxiv.1708.02182)
21. Tara N Sainath, Brian Kingsbury, Abdel-rahman Mohamed, George E Dahl, George Saon, Hagen Soltau, Tomas Beran, Aleksandr Y Aravkin, and Bhuvana Ramabhadran, “Improvements to deep convolutional neural networks for LVCSR,” in Automatic Speech Recognition and Understanding (ASRU), 2013 IEEE Workshop on. IEEE, 2013, pp. 315–320. [find](https://scholar.google.com/scholar?q=Tara+N+Sainath%2C+Brian+Kingsbury%2C+Abdel%2Drahman+Mohamed%2C+George+E+Dahl%2C+George+Saon%2C+Hagen+Soltau%2C+Tomas+Beran%2C+Aleksandr+Y+Aravkin%2C+and+Bhuvana+Ramabhadran%2C+%E2%80%9CImprovements+to+deep+convolutional+neural+n)

*Updated 24 October 2026*