Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CTC-DID: CTC-Based Arabic dialect identification for streaming applications

Muhammad Umar Farooq, Oscar Saz

TL;DR

This work reimagines Arabic dialect identification as an ASR-like, limited-vocabulary task by optimizing a CTC-based model that outputs dialect tags sequentially over time. Using a self-supervised encoder (mHuBERT) plus a transformer head, and training with repeated dialect labels via LAH or ASR-derived counts, the method enables streaming and robust performance on short utterances. The approach outperforms Whisper and ECAPA-TDNN on ADI-17 and achieves zero-shot superiority on Casablanca, demonstrating strong generalization under limited data. Its streaming-friendly design and potential for extension to code-switching and other ID tasks make it practically impactful for real-time dialect and related identification applications.

Abstract

This paper proposes a Dialect Identification (DID) approach inspired by the Connectionist Temporal Classification (CTC) loss function as used in Automatic Speech Recognition (ASR). CTC-DID frames the dialect identification task as a limited-vocabulary ASR system, where dialect tags are treated as a sequence of labels for a given utterance. For training, the repetition of dialect tags in transcriptions is estimated either using a proposed Language-Agnostic Heuristic (LAH) approach or a pre-trained ASR model. The method is evaluated on the low-resource Arabic Dialect Identification (ADI) task, with experimental results demonstrating that an SSL-based CTC-DID model, trained on a limited dataset, outperforms both fine-tuned Whisper and ECAPA-TDNN models. Notably, CTC-DID also surpasses these models in zero-shot evaluation on the Casablanca dataset. The proposed approach is found to be more robust to shorter utterances and is shown to be easily adaptable for streaming, real-time applications, with minimal performance degradation.

CTC-DID: CTC-Based Arabic dialect identification for streaming applications

TL;DR

This work reimagines Arabic dialect identification as an ASR-like, limited-vocabulary task by optimizing a CTC-based model that outputs dialect tags sequentially over time. Using a self-supervised encoder (mHuBERT) plus a transformer head, and training with repeated dialect labels via LAH or ASR-derived counts, the method enables streaming and robust performance on short utterances. The approach outperforms Whisper and ECAPA-TDNN on ADI-17 and achieves zero-shot superiority on Casablanca, demonstrating strong generalization under limited data. Its streaming-friendly design and potential for extension to code-switching and other ID tasks make it practically impactful for real-time dialect and related identification applications.

Abstract

This paper proposes a Dialect Identification (DID) approach inspired by the Connectionist Temporal Classification (CTC) loss function as used in Automatic Speech Recognition (ASR). CTC-DID frames the dialect identification task as a limited-vocabulary ASR system, where dialect tags are treated as a sequence of labels for a given utterance. For training, the repetition of dialect tags in transcriptions is estimated either using a proposed Language-Agnostic Heuristic (LAH) approach or a pre-trained ASR model. The method is evaluated on the low-resource Arabic Dialect Identification (ADI) task, with experimental results demonstrating that an SSL-based CTC-DID model, trained on a limited dataset, outperforms both fine-tuned Whisper and ECAPA-TDNN models. Notably, CTC-DID also surpasses these models in zero-shot evaluation on the Casablanca dataset. The proposed approach is found to be more robust to shorter utterances and is shown to be easily adaptable for streaming, real-time applications, with minimal performance degradation.
Paper Structure (11 sections, 3 figures, 2 tables, 1 algorithm)

This paper contains 11 sections, 3 figures, 2 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. CTC-based dialect identification system
  3. Streaming inference
  4. Experimental Setup
  5. Data
  6. Modelling
  7. Streaming
  8. Results and discussion
  9. Robustness to Short Utterances
  10. Streaming
  11. Conclusion

Figures (3)

  • Figure 1: Flow of the proposed dialect identification approach. The components within the shaded grey box represent modules that can either be frozen or fine-tuned during training. The green box highlights the inference pipeline.
  • Figure 2: The relationship between relative degradation in F1 score when only the utterances of length $\leq Duration$ are evaluated through the best model.
  • Figure 3: Effect on F1-score with different chunk sizes and context windows for streaming inference