---
title: "Automatic Long Audio Alignment and Confidence Scoring for"
description: "The authors collected about 1,400 hours of Al-Jazeerah talk shows together with their website transcripts. 13 references."
authors: ["Mohamed Elmahdy", "Mark Hasegawa-Johnson", "Eiman Mustafawi"]
year: 2014
venue: "Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC'14)"
source: "https://aclanthology.org/L14-1372/"
licence: "CC BY-NC 4.0 (ELRA, LREC proceedings)"
licence_url: "https://creativecommons.org/licenses/by-nc/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/elmahdy-2014-automatic-long-audio-alignment-confidence/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/elmahdy-2014-automatic-long-audio-alignment-confidence/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2014](https://md.research.elamur.ai/free/en/papers/2014/page/3/) / Automatic Long Audio Alignment and Confidence Scoring for Conversational Arabic Speech

# Automatic Long Audio Alignment and Confidence Scoring for Conversational Arabic Speech

**Authors:** Mohamed Elmahdy, [Mark Hasegawa-Johnson](https://md.research.elamur.ai/free/en/authors/mark-hasegawa-johnson/) and Eiman Mustafawi · **Year:** 2014 · **Venue:** [Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC'14)](https://md.research.elamur.ai/free/en/venues/lrec/) ([LREC 2014](https://md.research.elamur.ai/free/en/venues/lrec/2014/)) · **Publisher:** [European Language Resources Association (ELRA)](https://md.research.elamur.ai/free/en/publishers/european-language-resources-association-elra/) · **Licence:** [CC BY-NC 4.0 (ELRA, LREC proceedings)](https://creativecommons.org/licenses/by-nc/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/L14-1372/) · [PDF](https://aclanthology.org/L14-1372.pdf) · [OpenAlex](https://openalex.org/W2251459480)

**Citations:** 2 citations (OpenAlex, 2026-10-14)

## Summary
The paper addresses long audio alignment for conversational Arabic: synchronizing hour-scale broadcast recordings with human transcripts that carry no timing. This is harder than in English because the speech is spontaneous and largely dialectal. The authors collected about 1,400 hours of Al-Jazeerah talk shows together with their website transcripts. Their pipeline first splits each episode into 5–10 s chunks using energy-based silence detection, then merges the chunks. Alignment runs in two recognition passes with a CMU Sphinx GMM-HMM recognizer.

- **Pass 1** uses a bigram LM biased toward the episode's own text, a graphemic lexicon, and a standard-Arabic acoustic model.
- **Pass 2** adds unsupervised MLLR adaptation and a small LM per segment built from the neighbouring aligned text.

Levenshtein matching maps recognizer output onto the transcript. On a disfluent conversational test set, word alignment accuracy rose from 95.4% to 97.7%. A segment-level confidence score then lets low-quality segments be discarded, which lifts accuracy to about 99% or higher. The approach makes large untimed archives usable for training acoustic models.

## Key points
- **Data:** about 2,150 Al-Jazeerah episodes (20–50 min each, 2008–2012), totalling more than 1,400 hours. Audio was taken from YouTube and resampled to 16 kHz mono, and transcripts were scraped from the Al-Jazeerah website with rule-based cleanup.
- **Evaluation sets:** conv-eval-set (5 h of manually aligned conversational podcasts) and read-eval-set (5 h from the LDC Arabic Broadcast News corpus, with timing and disfluency tags removed).
- **Segmentation:** energy is computed over 512-sample windows, and the silence threshold is 20% of the episode's mean energy. A silence must last at least 350 ms, so geminates and quiet sounds such as /s/ do not trigger a cut. Cuts fall at the centre of each silence, and short pieces are then merged up to 5–10 s.
- **Acoustic modelling:** a graphemic AM was trained on more than 60 h of standard Arabic, with 39 MFCC features reduced to 29 dimensions via LDA/MLLT. Pass 1 uses 8 Gaussians per state and pass 2 uses 16.
- **Results on conv-eval-set:** anchor rate rose from 84.6% to 87.1% to 91.9%, and word accuracy from 95.4% to 96.2% to 97.7%. On read-eval-set the baseline is 99.0%. Keeping only segments with confidence above 90% gives 99.8% while dropping 10% of segments. On conv-eval-set, thresholding reaches 99.2%.

## Interesting ideas
- Matching recognizer output to the reference with Levenshtein alignment, rather than trusting it sequentially, contains each error locally, so one bad segment cannot shift the alignment of everything after it.
- The anchor rate (share of reference words exactly recovered by the system) works as a cheap, label-free signal that tracks the true downstream accuracy.
- Reporting a curve of accuracy against the fraction of rejected items, across several confidence thresholds, makes the cost of filtering explicit instead of hiding it behind a single number.
- After a coarse first pass, each unit is re-decoded with a tiny model built only from its own text plus its neighbours, which fixes the boundary errors that dominate the residue.
- Using spelling as the pronunciation (graphemic lexicon) sidesteps an unmaintainable variant dictionary for non-standard language, letting the model absorb variation implicitly.

## For Mova

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

- **I1** (tokens): align tokenizer/normalizer output back to raw character offsets with edit-distance matching, so each token stays traceable to its source span even after a local mis-step.
- **I2** (learn): measure how many dependency arcs or lemmas several witness parsers agree on per sentence, and use that agreement rate as an unsupervised quality gauge for silver data.
- **I3** (learn): when filtering silver treebank sentences by confidence, publish accuracy against the share of data discarded, so threshold choices are explainable and the data sacrificed per gain is visible.
- **I4** (back): when testing the text → graph → text compressor, align regenerated text to the original and re-examine only mismatched spans, using just the surrounding sentences as context, to localize which graph fragment lost information.
- **I5** (tokens): for informal or misspelled English, fall back to surface-form features instead of hand-listing variant spellings in the lexicon, and flag such fallbacks explicitly so the step stays explainable.

*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

- [University of Illinois Urbana-Champaign](https://md.research.elamur.ai/free/en/universities/university-of-illinois-urbana-champaign/) — University, Urbana, United States · 108 works · 67,432 citations
- [Qatar University](https://md.research.elamur.ai/free/en/universities/qatar-university/) — University, Doha, Qatar · 10 works · 81 citations

📚 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 (3): [🗨️ Dialogue systems](https://md.research.elamur.ai/free/en/topics/dialogue-systems/) · 401 works · [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/) · 1,484 works · [🎙️ Speech & spoken language](https://md.research.elamur.ai/free/en/topics/speech/) · 233 works

## Cite

Cite as: Mohamed Elmahdy, Mark Hasegawa-Johnson and Eiman Mustafawi. 2014. Automatic Long Audio Alignment and Confidence Scoring for Conversational Arabic Speech. Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC'14).

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

```bibtex
@inproceedings{elmahdy2014automatic,
  title = {{Automatic Long Audio Alignment and Confidence Scoring for Conversational Arabic Speech}},
  author = {Mohamed Elmahdy, Mark Hasegawa-Johnson and Eiman Mustafawi},
  booktitle = {Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC'14)},
  year = {2014},
  doi = {10.63317/3udpbscws4rw},
  url = {https://doi.org/10.63317/3udpbscws4rw}
}
```

- [← previous Assessing the so called Marked Inflectional Features of the Nigerian English: a Second Language Acquisition Theory Account](https://md.research.elamur.ai/free/en/papers/oshodi-2014-assessing-so-called-marked-inflectional/)
- [next → Machine Learning based English Sentiment Analysis](https://md.research.elamur.ai/free/en/papers/tran-2019-machine-learning-based-english-sentiment/)
- [↑ back to papers of 2014](https://md.research.elamur.ai/free/en/papers/2014/page/3/)

## References

13 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Billa, J., Noamany, M., Srivastava, A., Liu, D., Stone, R., Xu, J., Makhoul, J., and Kubala, F. (2002). Audio indexing of Arabic broadcast news. In Proceedings of ICASSP, volume 1, pages 5–8. [find](https://scholar.google.com/scholar?q=Billa%2C+J%2E%2C+Noamany%2C+M%2E%2C+Srivastava%2C+A%2E%2C+Liu%2C+D%2E%2C+Stone%2C+R%2E%2C+Xu%2C+J%2E%2C+Makhoul%2C+J%2E%2C+and+Kubala%2C+F%2E+%282002%29%2E+Audio+indexing+of+Arabic+broadcast+news%2E+In+Proceedings+of+ICASSP%2C+volume+1%2C+pages+5%E2%80%938%2E)
2. Elmahdy, M., Gruhn, R., Minker, W., and Abdennadher, S. (2009). Survey on common Arabic language forms from a speech recognition point of view. In Proceedings of the International Conference on Acoustics (NAG/DAGA), pages 63–66. [find](https://scholar.google.com/scholar?q=Elmahdy%2C+M%2E%2C+Gruhn%2C+R%2E%2C+Minker%2C+W%2E%2C+and+Abdennadher%2C+S%2E+%282009%29%2E+Survey+on+common+Arabic+language+forms+from+a+speech+recognition+point+of+view%2E+In+Proceedings+of+the+International+Conference+on+Acoust)
3. Elmahdy, M., Gruhn, R., Minker, W., and Abdennadher, S. (2010). Cross-lingual acoustic modeling for dialectal Arabic speech recognition. In Proceedings of INTER-SPEECH, pages 873–876. [find](https://scholar.google.com/scholar?q=Elmahdy%2C+M%2E%2C+Gruhn%2C+R%2E%2C+Minker%2C+W%2E%2C+and+Abdennadher%2C+S%2E+%282010%29%2E+Cross%2Dlingual+acoustic+modeling+for+dialectal+Arabic+speech+recognition%2E+In+Proceedings+of+INTER%2DSPEECH%2C+pages+873%E2%80%93876%2E)
4. Elmahdy, M., Gruhn, R., and Minker, W. (2011). Novel Techniques for Dialectal Arabic Speech Recognition. Springer, New York, first edition. [↗](https://doi.org/10.1007/978-1-4614-1906-8)
5. Elmahdy, M., Hasegawa-Johnson, M., and Mustafawi, E. (2013). A framework for conversational arabic speech long audio alignment. In Proceedings of the 6th Language and Technology Conference (LTC), pages 290– 293. [find](https://scholar.google.com/scholar?q=Elmahdy%2C+M%2E%2C+Hasegawa%2DJohnson%2C+M%2E%2C+and+Mustafawi%2C+E%2E+%282013%29%2E+A+framework+for+conversational+arabic+speech+long+audio+alignment%2E+In+Proceedings+of+the+6th+Language+and+Technology+Conference+%28LTC%29%2C+page)
6. Ferguson, C. A. (1959). Diglossia. Word, 15:325–340. [find](https://scholar.google.com/scholar?q=Ferguson%2C+C%2E+A%2E+%281959%29%2E+Diglossia%2E+Word%2C+15%3A325%E2%80%93340%2E)
7. Hazen, T.J. (2006). Automatic alignment and error correction of human generated transcripts for long speech recordings. In Proceedings of INTERSPEECH, pages 1606–1609. [find](https://scholar.google.com/scholar?q=Hazen%2C+T%2EJ%2E+%282006%29%2E+Automatic+alignment+and+error+correction+of+human+generated+transcripts+for+long+speech+recordings%2E+In+Proceedings+of+INTERSPEECH%2C+pages+1606%E2%80%931609%2E)
8. Huggins-Daines, D., Kumar, M., Chan, A., Black, A.W., Ravishankar, M., and Rudnicky, A.I. (2006). Pocketsphinx: A free, real-time continuous speech recognition system for hand-held devices. In Proceedings of ICASSP, volume 1, pages 185–188. [find](https://scholar.google.com/scholar?q=Huggins%2DDaines%2C+D%2E%2C+Kumar%2C+M%2E%2C+Chan%2C+A%2E%2C+Black%2C+A%2EW%2E%2C+Ravishankar%2C+M%2E%2C+and+Rudnicky%2C+A%2EI%2E+%282006%29%2E+Pocketsphinx%3A+A+free%2C+real%2Dtime+continuous+speech+recognition+system+for+hand%2Dheld+devices%2E+In+Proceed)
9. Leggetter, C.J. and Woodland, P.C. (1995). Maximum likelihood linear regression for speaker adaptation of the parameters of continuous density hidden Markov models. Computer Speech and Language, 9:171–185. [find](https://scholar.google.com/scholar?q=Leggetter%2C+C%2EJ%2E+and+Woodland%2C+P%2EC%2E+%281995%29%2E+Maximum+likelihood+linear+regression+for+speaker+adaptation+of+the+parameters+of+continuous+density+hidden+Markov+models%2E+Computer+Speech+and+Language%2C+9%3A171)
10. Liu, Y., Shriberg, E., and Stolcke, A. Y. (2003). Automatic disfluency identification in conversational speech using multiple knowledge sources. In Proceedings of Eurospeech, pages 957–960. [find](https://scholar.google.com/scholar?q=Liu%2C+Y%2E%2C+Shriberg%2C+E%2E%2C+and+Stolcke%2C+A%2E+Y%2E+%282003%29%2E+Automatic+disfluency+identification+in+conversational+speech+using+multiple+knowledge+sources%2E+In+Proceedings+of+Eurospeech%2C+pages+957%E2%80%93960%2E)
11. Maamouri, M., Graff, D., and Cieri, C. (2006a). Arabic Broadcast News Speech, LDC Catalog No.: LDC2006S46. Linguistic Data Consortium, Philadelphia. [find](https://scholar.google.com/scholar?q=Maamouri%2C+M%2E%2C+Graff%2C+D%2E%2C+and+Cieri%2C+C%2E+%282006a%29%2E+Arabic+Broadcast+News+Speech%2C+LDC+Catalog+No%2E%3A+LDC2006S46%2E+Linguistic+Data+Consortium%2C+Philadelphia%2E)
12. Maamouri, M., Graff, D., and Cieri, C. (2006b). Arabic Broadcast News Transcripts, LDC Catalog No.: LDC2006T20. Linguistic Data Consortium, Philadelphia. Moreno, P.J., Joerg, C., Thong, JM. [find](https://scholar.google.com/scholar?q=Maamouri%2C+M%2E%2C+Graff%2C+D%2E%2C+and+Cieri%2C+C%2E+%282006b%29%2E+Arabic+Broadcast+News+Transcripts%2C+LDC+Catalog+No%2E%3A+LDC2006T20%2E+Linguistic+Data+Consortium%2C+Philadelphia%2E+Moreno%2C+P%2EJ%2E%2C+Joerg%2C+C%2E%2C+Thong%2C+JM%2E)
13. V., and Glickman, O. (1998). A recursive algorithm for the forced alignment of very long audio segments. In Proceedings of ICSLP. [find](https://scholar.google.com/scholar?q=V%2E%2C+and+Glickman%2C+O%2E+%281998%29%2E+A+recursive+algorithm+for+the+forced+alignment+of+very+long+audio+segments%2E+In+Proceedings+of+ICSLP%2E)

*Updated 24 October 2026*