---
title: "Advances in the CMU/Interact Arabic GALE Transcription System"
description: "The paper reports nine months of work on the CMU/InterACT Arabic speech recognizer for… 18.3% WER. 10 references, cited by 1 paper here."
authors: ["Mohamed Noamany", "Thomas Schaaf", "Tanja Schultz"]
year: 2007
venue: "Human Language Technologies 2007: The Conference of the North American Chapter of the Association for Computational Linguistics; Companion Volume, Short Papers"
source: "https://aclanthology.org/N07-2033/"
licence: "CC BY-NC-SA 3.0 (ACL Anthology, before 2016)"
licence_url: "https://creativecommons.org/licenses/by-nc-sa/3.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/noamany-2007-advances-cmu-interact-arabic-gale/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/noamany-2007-advances-cmu-interact-arabic-gale/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2007](https://md.research.elamur.ai/free/en/papers/2007/page/2/) / Advances in the CMU/Interact Arabic GALE Transcription System

# Advances in the CMU/Interact Arabic GALE Transcription System

**Authors:** Mohamed Noamany, Thomas Schaaf and [Tanja Schultz](https://md.research.elamur.ai/free/en/authors/tanja-schultz/) · **Year:** 2007 · **Venue:** [Human Language Technologies 2007: The Conference of the North American Chapter of the Association for Computational Linguistics; Companion Volume, Short Papers](https://md.research.elamur.ai/free/en/venues/naacl/) ([NAACL 2007](https://md.research.elamur.ai/free/en/venues/naacl/2007/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY-NC-SA 3.0 (ACL Anthology, before 2016)](https://creativecommons.org/licenses/by-nc-sa/3.0/)

**Links:** [Publisher's page](https://aclanthology.org/N07-2033/) · [PDF](https://aclanthology.org/N07-2033.pdf) · [OpenAlex](https://openalex.org/W2064203761)

**Citations:** 11 citations (OpenAlex, 2026-10-14) · 1 in our library, of the 149,439 papers whose bibliographies we read

## Summary
The paper reports nine months of work on the CMU/InterACT Arabic speech recognizer for broadcast news and broadcast conversation, built for the GALE 2006 evaluation. Two problems drive the work. Written Modern Standard Arabic usually omits short vowels, and there is little conversational training audio. The system runs on JRTk with the IBIS decoder. The authors improved it in several ways:
- They built a vowelized pronunciation lexicon using the Buckwalter morphological analyzer and the LDC Arabic Tree-Bank.
- They cross-adapted the vowelized system with a non-vowelized grapheme-based one.
- They aligned untimed Al-jazeera web transcripts with recorded shows to create extra training data with little supervision.
- They interpolated 11 source-specific 4-gram language models.
- They tuned acoustic model size and added MMIE training.

Taken together, these steps cut word error rate on RT04 by about 40% relative to the legacy system, reaching 18.3% WER. The work shows that free, imperfect web text can be turned into usable training data. It also shows that two lexical views of the same language can complement each other.

## Key points
- Acoustic training used 190 hours, drawn from FBIS, LDC-TDT4, Al-jazeera BC recordings and LDC-GALE. Out of 80 hours of rough TDT4 transcription, 50 were kept, which gave a 12% relative gain.
- Web transcripts were aligned with recorded Al-jazeera shows, and only matching stretches of at least 3 consecutive words were kept. This yielded about 30 usable hours and a 7% relative improvement on RT04.
- Vowelization covered a 129k-word vocabulary:
  - The Buckwalter analyzer vowelized 106K entries.
  - The Tree-Bank added 5k more, and 11k stayed unresolved.
  - The resulting 559,035 pronunciations were cut by about 27% to 407,754 through forced alignment.
- WER on RT04 (speaker-adaptive pass) dropped in three stages:
  - The non-vowelized system scored 20.8%.
  - The vowelized system scored 19.3%.
  - A 3-pass cross-adapted combination (vowelized SI → non-vowelized SA → vowelized SA) reached 18.3%.
- Other results:
  - Merging glottal-stop letter variants at the start of words lowered LM perplexity by 20% and improved WER by 0.9%.
  - The final language model, built with SRILM, has 126M n-grams, a 219k vocabulary and perplexity 212.
  - Decoding speed doubled from 20RT to 10RT at a cost of only 0.2% WER.

## Interesting ideas
- To get training data from transcripts that have no timing, decode the audio with a model biased toward that transcript and keep only runs where at least three words in a row agree; this turns noisy free text into trusted supervision cheaply.
- Look up each word's analysis through a fixed chain of sources: morphological analyzer first, then treebank, and the raw form as a last resort. Counting how many words each source resolves keeps coverage gaps visible.
- When a generator overproduces variants, remove those never seen in aligned real data; a quarter of the pronunciations went away without harm.
- Two systems that differ in only one layer (here the lexicon and decision tree), while sharing vocabulary and LM, make good complementary partners; feeding one's output in to adapt the other beat either alone, and the order of passes mattered.
- Normalize spelling variants that are often written inconsistently, aligned with the conventions of the downstream consumer (here translation); this cut perplexity sharply.

## For Mova

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

- **I1** (learn): When building silver UD data, accept only spans where independent witness parsers agree over several consecutive tokens, instead of whole sentences or single arcs.
- **I2** (tokens): Lemma and morphology lookup in crate en could record which source answered (lexicon, treebank, fallback), making every lemma explainable and the unresolved set easy to target.
- **I3** (tokens): Candidate morphological analyses or tag sets that never occur in silver-parsed corpora could be pruned, shrinking the tagger's search space.
- **I4** (tree): The perceptron parser and gparse could be chained so one's output conditions the other's second pass, with the order chosen empirically rather than assumed.
- **I5** (tokens): Fold quote, apostrophe and hyphen variants into canonical forms that match what the frames and world stages expect, logging each mapping 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

- [Carnegie Mellon University](https://md.research.elamur.ai/free/en/universities/carnegie-mellon-university/) — University, Pittsburgh, United States · 539 works · 179,533 citations

📚 Shelves (5): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 16,667 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 16,194 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 16,039 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 15,891 works · [🔊 Signal Processing](https://md.research.elamur.ai/free/en/shelves/computer-science/signal-processing/) · 452 works

🧭 Topics (1): [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/) · 2,908 works

## Cite

Cite as: Mohamed Noamany, Thomas Schaaf and Tanja Schultz. 2007. Advances in the CMU/Interact Arabic GALE Transcription System. Human Language Technologies 2007: The Conference of the North American Chapter of the Association for Computational Linguistics; Companion Volume, Short Papers.

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

```bibtex
@inproceedings{noamany2007advances,
  title = {{Advances in the CMU/Interact Arabic GALE Transcription System}},
  author = {Mohamed Noamany, Thomas Schaaf and Tanja Schultz},
  booktitle = {Human Language Technologies 2007: The Conference of the North American Chapter of the Association for Computational Linguistics; Companion Volume, Short Papers},
  year = {2007},
  doi = {10.3115/1614108.1614141},
  url = {https://doi.org/10.3115/1614108.1614141}
}
```

- [← previous SVM Model Tampering and Anchored Learning: A Case Study in Hebrew NP Chunking](https://md.research.elamur.ai/free/en/papers/goldberg-2007-svm-model-tampering-anchored-learning/)
- [next → HIT-IR-WSD: A WSD System for English Lexical Sample Task](https://md.research.elamur.ai/free/en/papers/guo-2007-hit-ir-wsd-wsd-system/)
- [↑ back to papers of 2007](https://md.research.elamur.ai/free/en/papers/2007/page/2/)

## References

10 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. J. Billa, M. Noamany, A. Srivastava, D. Liu, R. Stone, J. Xu, J. Makhoul, and F. Kubala, “Audio Indexing of Arabic Broadcast News”, International Conference on Acoustics, Speech, and Signal Processing (ICASSP), May 2002. [find](https://scholar.google.com/scholar?q=J%2E+Billa%2C+M%2E+Noamany%2C+A%2E+Srivastava%2C+D%2E+Liu%2C+R%2E+Stone%2C+J%2E+Xu%2C+J%2E+Makhoul%2C+and+F%2E+Kubala%2C+%E2%80%9CAudio+Indexing+of+Arabic+Broadcast+News%E2%80%9D%2C+International+Conference+on+Acoustics%2C+Speech%2C+and+Signal+Processing)
2. H. Yu, Y-C. Tam, T. Schaaf, S. Stüker, Q. Jin, M. Noamany, and T. Schultz “The ISL RT04 Mandarin Broadcast News Evaluation System”, EARS Rich Transcription workshop, Palisades, NY, 2004. [find](https://scholar.google.com/scholar?q=H%2E+Yu%2C+Y%2DC%2E+Tam%2C+T%2E+Schaaf%2C+S%2E+St%C3%BCker%2C+Q%2E+Jin%2C+M%2E+Noamany%2C+and+T%2E+Schultz+%E2%80%9CThe+ISL+RT04+Mandarin+Broadcast+News+Evaluation+System%E2%80%9D%2C+EARS+Rich+Transcription+workshop%2C+Palisades%2C+NY%2C+2004%2E)
3. L. Nguyen et al., “Light supervision in acoustic model training,” ICASSP , Montreal, QC, Canada, May 2004. [find](https://scholar.google.com/scholar?q=L%2E+Nguyen+et+al%2E%2C+%E2%80%9CLight+supervision+in+acoustic+model+training%2C%E2%80%9D+ICASSP+%2C+Montreal%2C+QC%2C+Canada%2C+May+2004%2E)
4. M. Afify et al.,"Recent progress in Arabic broadcast news transcription at BBN", In INTERSPEECH-2005. [find](https://scholar.google.com/scholar?q=M%2E+Afify+et+al%2E%2C%22Recent+progress+in+Arabic+broadcast+news+transcription+at+BBN%22%2C+In+INTERSPEECH%2D2005%2E)
5. A. Stolcke SRILM- An Extensible Language Modeling ToolKit ICSLP. 2002, Denver, Colorado. [find](https://scholar.google.com/scholar?q=A%2E+Stolcke+SRILM%2D+An+Extensible+Language+Modeling+ToolKit+ICSLP%2E+2002%2C+Denver%2C+Colorado%2E)
6. T.Buckwalter, “[Issues in Arabic Orthography and morphology Analysis](https://md.research.elamur.ai/free/en/papers/buckwalter-2004-issues-arabic-orthography-morphology-analysis/)”, COLING 2004, Geneva, 2004.
7. Q. Jin, T. Schultz, “Speaker segmentation and clustering in meetings”, ICSLP, 2004. [find](https://scholar.google.com/scholar?q=Q%2E+Jin%2C+T%2E+Schultz%2C+%E2%80%9CSpeaker+segmentation+and+clustering+in+meetings%E2%80%9D%2C+ICSLP%2C+2004%2E)
8. P. Zhan, M. Westphal, "Speaker Normalization Based On Frequency Warping", ICASSP 1997, Munich, Germany. [↗](https://doi.org/10.5445/ir/309097)
9. M. Finke, et al., "The Karlsruhe Verbmobil Speech Recognition Engine," International Conference on Acoustics, Speech, and Signal Processing, ICASSP,1997. [find](https://scholar.google.com/scholar?q=M%2E+Finke%2C+et+al%2E%2C+%22The+Karlsruhe+Verbmobil+Speech+Recognition+Engine%2C%22+International+Conference+on+Acoustics%2C+Speech%2C+and+Signal+Processing%2C+ICASSP%2C1997%2E)
10. H. Soltau, et al., "A One Pass-Decoder Based On Polymorphic Linguistic Context", ASRU 2001, Trento, Italy, 2001 [find](https://scholar.google.com/scholar?q=H%2E+Soltau%2C+et+al%2E%2C+%22A+One+Pass%2DDecoder+Based+On+Polymorphic+Linguistic+Context%22%2C+ASRU+2001%2C+Trento%2C+Italy%2C+2001)

*Updated 1 November 2026*