📄 Development of a TV Broadcasts Speech Recognition System for Qatari Arabic

Summary

The paper tackles speech recognition for Qatari Arabic (QA), a Gulf sub-dialect with very little transcribed audio. Formal MSA, by contrast, is well resourced. The authors collected and transcribed a 15-hour wide-band QA corpus from Qatari TV series and talk shows. Transcription uses ordinary Arabic script plus five Persian letters for sounds MSA lacks. On this corpus they built a Kaldi GMM-HMM baseline with grapheme-based acoustic units. They then transferred knowledge from MSA in several steps:

• orthographic normalization of hypotheses and references before scoring;
• interpolation of the small QA trigram LM with a large MSA LM trained on Arabic Gigaword;
• phone mapping and pooling of MSA and QA audio;
• MLLR and MAP adaptation of the pooled model toward QA;
• system combination with lattice MBR decoding.

Together these steps cut WER by 21.3% relative on the development set and 28.9% on the evaluation set. The main lesson is that a large standard-variety resource can help a sparse dialect even when the two differ substantially. The key condition is that the small dialect set is re-weighted afterwards so it is not drowned out.

Key points

• QA corpus (publicly downloadable): 15 h in total (Tesaneef 9.3 h, Sabah El-Doha 2.0 h, Al-Jazeerah 3.7 h), split into 13 h train, 1 h dev and 2 h eval.

• MSA audio: about 62.4 h of broadcast news from the ELRA NEMLAR and NetDC corpora.

• Baseline system:

  • Kaldi GMM-HMM with MMIE training, LDA+MLLT and fMLLR SAT.
  • 41 grapheme units, 1000 tied states, 8 Gaussians per state.
  • QA-only trigram LM with a vocabulary of 15.5K words.
  • WER of 61.7% on dev and 80.8% on eval.

• LM interpolation: combining the QA LM with an MSA LM (Gigaword, 800M+ words, top 256K vocabulary) gives a 265.7K vocabulary. OOV drops from 18.4% to 8.9% on dev and from 22.2% to 9.2% on eval.

• WER progression (dev / eval):

  Stage	Dev WER	Eval WER
  After orthographic normalization	60.9%	79.9%
  Cross-lingual LM	56.0%	64.4%
  Pooling + MLLR/MAP adaptation	55.6%	62.5%
  Lattice MBR	47.9%	56.8%

Interesting ideas

• I1. When a language has no fixed spelling, the authors map harmless orthographic variants (Alef forms, Teh Marbuta, final Yeh) to one form in both the system output and the gold reference before alignment, so the metric counts only real errors.
• I2. Training first on pooled big-plus-small data and then adapting toward the small set counteracts the large source drowning out the target's features; MAP adaptation explicitly raises the effective weight of target samples.
• I3. Linearly mixing a tiny in-domain model with a huge out-of-domain one, with weights tuned only on a dev set, roughly halved OOV while keeping dialect-specific statistics.
• I4. Rather than romanizing dialect speech, the authors extended the native script with a few extra letters for sounds the standard script lacks, which keeps transcripts compatible with standard-language resources.
• I5. Combining hypotheses through minimum-Bayes-risk decoding gave the largest single gain, by choosing the output with the lowest expected error rather than the single top-scoring path.

For Mova

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

• I1 (learn): Mova's evaluation could apply a declared, logged normalization to both sides (spelling variants, quote styles, hyphenation) before scoring, keeping a separate un-normalized score as a control.
• I2 (tree): the perceptron parser or gparse could be trained on pooled UD treebanks and then fine-tuned with up-weighted target-genre trees (e.g., fables or school texts) so genre-specific constructions are not swamped.
• I3 (tokens): lemma and morphology lexicons could be built as a dev-tuned blend of a small domain lexicon and a broad general one, with each lookup recording which source supplied the answer for explainability.
• I4 (tokens): Mova could mark nonstandard or dialectal forms (eye-dialect, archaic spellings in tales) with explicit tags while keeping them close to standard spelling, so the standard lexicon still applies.
• I5 (learn): when several witness parsers disagree on silver data, Mova could pick the tree that minimizes expected disagreement with all witnesses (an MBR consensus) instead of using plain majority voting per arc.