Automatic Dialect Detection in Arabic Broadcast Speech

1. Introduction

• A good DID system can identify dialectal segments in mixed, untranscribed speech, helping reduce dialectal ASR word error rates through dialect-specific training or model adaptation.The paper describes DID as a special case of language identification and notes growing interest in the task within the ASR community.Source: §1 · p. 1
• Arabic dialects are grouped into Egyptian, North African, Gulf, Levantine, and Modern Standard Arabic; lexical variation across them motivates dialect identification.Table 1 provides lexical examples in Arabic and Buckwalter format.Source: §1 · p. 1, Table 1 · p. 1
• Prior language-identification work spans acoustic features and phonetic or lexical features, including words, roots, morphology, grammars, and phone posteriors.The cited approaches also include i-vectors, SVMs, non-negative factor analysis, and phoneme-sequence models for Arabic DID and LID.Source: §1 · p. 1
• The paper investigates lexical, phonetic, and i-vector Vector Subspace Models, studies their feature selection and interaction, and releases an Arabic DID system for further research.The authors describe the released system as a resource others can extend to improve DID performance.Source: §1 · p. 1

2.1. Senone based Utterance VSM

• Senone-based utterance vectors count phone n-grams extracted from phoneme-recognizer output, with identity or tf.idf scaling; the vector space has d senone dimensions across N utterances.Senones are n-gram phone sequences with n ≤4; the phoneme sequence is derived through automatic vowelization and vowelization-to-phonetization.Source: §2.1 · pp. 1–2, Figure 1 · p. 2

2.2. Word based Utterance VSM

• Word-based utterance vectors count ASR-extracted words with a scaling function; the reported vocabulary has 55k words and the tri-gram dictionary has 580k entries.The scaling function has the same interpretation as for the senone-based space.Source: §2.2 · p. 2

2.3.1. Bottleneck Features (BN)

• The bottleneck-feature setup uses two successive DNNs, each with five hidden sigmoid layers and one linear bottleneck layer, with tied-state targets.The first network takes stacked Mel-filter-bank energies with pitch and voicing probability; the second uses first-network bottleneck outputs and five frame offsets, spanning 31 frames.Source: §2.3.1 · p. 2

2.3.2. Modeling

• The i-vector framework represents an utterance through a low-dimensional factor for its mean offset from a channel- and dialect-independent supervector; experiments use 2048-component UBM GMMs and 400-dimensional i-vectors.The factors are treated as latent variables, with the i-vector as their maximum-a-posteriori point estimate; LDA and within-class covariance normalization are used for discrimination.Source: §2.3.2 · p. 2

3.1. Train Data

• Training recordings come from broadcast news in Egyptian, Levantine, Gulf, North African, and Modern Standard Arabic, were recorded at 16 kHz, and were segmented to remove overlap and non-speech audio.Table 2 reports available speech hours by dialect.Source: §3.1 · p. 2, Table 2 · p. 3

3.2. Test Data

• The test set was crowd-labeled with at least three and up to nine judges per file, or until 75% inter-annotator agreement was reached.The test recordings came from the same broadcast domain but used different channels and recording setup, reducing sensitivity to channel and speaker characteristics.Source: §3.2 · p. 3
• The paper provides the train and test data on the QCRI web portal; Table 3 reports test utterance counts by dialect.The test set differs from earlier data in its reproducibility and recording conditions.Source: §3.2 · p. 3, Table 3 · p. 3

4.1. Choosing the Best Classifier

• Linear SVM performed best among the tested classifiers and was selected for the remaining experiments.The comparison used lexical utterance vectors with a 55k lexicon and no dimensionality reduction.Source: §4.1 · p. 3

4.2. Feature Selection Study

• tf.idf scaling significantly improved accuracy over identity scaling for lexical utterance vectors.Source: §4.2 · p. 3
• Senone-based spaces were more discriminative than word-based spaces, while tf.idf did not help senone features.Source: §4.2 · p. 3
• Combining the best senone-based and lexical-based spaces did not help, which the authors attribute to their capturing similar dialect information.Source: §4.2 · p. 3
• The 400-dimensional bottleneck-feature i-vector space was significantly more discriminative than previously defined spaces, and LDA plus WCNN significantly improved DID results.Source: §4.2 · p. 3
• Concatenating the best senone-based and i-vector-based utterance vectors produced slight improvements.Source: §4.2 · p. 3

4.3. One Vs All classification (Sanity Check)

• 100% accuracy was achieved for both English-versus-Arabic and MSA-versus-dialectal Arabic binary classification.Source: §4.3 · p. 4

4.4. System Output Combination

• 60.2% accuracy was the best reported result, from fusing the best senone system with the SVM-based i-vector system.The authors offer the systems’ different error patterns as one explanation for the gain.Source: §4.4 · p. 4

5. Discussion

• Gulf and Levantine are the most confusable dialect pair in the best DID system.The authors believe this relates to greater lexical similarity between the dialects.Source: Table 7 · p. 4, §5 · p. 4
• More than 50% of sampled segments from potential Gulf speakers were MSA speech, indicating within-speaker switching between dialect and MSA.Table 8 concerns the expected dialect of speech segments from particular dialectal speakers.Source: §5 · p. 4, Table 8 · p. 4

6. Conclusions

• 60.2% accuracy was achieved for dialect classification using system combination, alongside 100% accuracy on MSA versus Dialectal Arabic and English versus Arabic.The paper reports the 60.2% result for system combination and 100% for each of the two binary classification tasks.Source: §6 · p. 4
• The authors studied potential code-switching patterns in the classifier and their correlation with manual annotation.Source: §6 · p. 4
• Future work includes studying MSA–dialect switching without speaker diarization or silence gaps and using deep neural classifiers for nonlinear decision boundaries.Source: §6 · p. 4

• “bnf.png” is available in PNG format.Source: http://arxiv.org/ps/1509.06928v2 · p. 7