Addressing Code-Switching in French/Algerian Arabic Speech

Djegdjiga Amazouz, Martine Adda-Decker, Lori LamelInterspeech 2017

TL;DR

This paper tackles the challenge of automatically analyzing code-switching (CS) in French/Algerian Arabic speech, a common phenomenon in bilingual communities. They introduce a new, high-quality corpus of French/Algerian Arabic CS speech and explore how speech technologies like language identification and automatic speech recognition can be used to detect and quantify this linguistic behavior.

What Problem Does This Paper Address?

Code-switching, the practice of alternating between two or more languages within a single conversation, is a widespread phenomenon among bilingual and multilingual individuals. While linguists have studied CS for decades, its integration into automatic speech processing (ASR) systems remains a significant challenge. Existing research often relies on broadcast or web data, which can suffer from low acoustic quality and limited CS density. This paper aims to bridge this gap by creating a high-quality, annotated corpus of French/Algerian Arabic CS speech and investigating how speech technologies can be leveraged to analyze and understand this complex linguistic behavior. Specifically, they want to know if CS is common in Maghrebian broadcast speech, how to collect such data, and if current speech technologies can effectively detect and quantify CS.

Methods

The authors employed a multi-pronged approach:

  • Corpus Collection: They designed a specific protocol to collect a French/Algerian Arabic CS speech corpus. This involved inviting bilingual speakers to a lab setting and creating a sociolinguistic context that encouraged natural code-switching. The protocol included controlled read speech tasks (to assess monolingual capabilities) and dual conversations with a linguist who actively triggered CS through targeted questions.
  • Linguistic Description: They provided a linguistic description of the collected CS stretches, categorizing them into intra-sentential and inter-sentential switches and analyzing speech duration in each language.
  • Speech Technology Application: They explored the use of Automatic Speech Recognition (ASR) systems for detecting CS. Specifically, they used word posteriors from parallel French and Arabic ASR systems to analyze language identification and the potential for detecting language shifts within speech segments.

Datasets

  • MCSM Corpus: A preliminary study used a large collection of broadcast and podcast speech data from Maghrebian countries (Algeria, Morocco, Tunisia) to assess the prevalence of Arabic/French CS.
  • FACST Corpus (French Arabic Code-switching Triggered): This is a newly introduced benchmark corpus specifically designed for this study. It consists of approximately 4 hours of transcribed speech from 8 bilingual French/Algerian Arabic speakers. The corpus includes both read speech and spontaneous conversations, with a focus on high acoustic quality and a relatively high density of CS.

Key Insights and Ideas

  • Algerian Arabic Leads in CS: The preliminary study on the MCSM corpus revealed that Algerian Arabic/French CS is significantly more frequent in broadcast media compared to Moroccan and Tunisian Arabic/French CS.
  • Controlled Elicitation is Key: The FACST corpus demonstrates that a carefully designed lab-based protocol, involving a linguist actively triggering CS, can successfully elicit naturalistic code-switching phenomena, even in a controlled environment.
  • Short CS Segments are Challenging: The paper highlights that short-duration CS segments pose significant challenges for automatic language identification.
  • Parallel ASR Systems Show Promise: Despite the challenges, parallel ASR systems (one for French, one for Algerian Arabic) can provide word posteriors that act as good indicators of language change, offering a potential avenue for CS detection.
  • Speech Technologies Aid Annotation: Automatic speech transcription and temporal alignment tools are invaluable for creating high-quality transcriptions and facilitating large-scale acoustic-phonetic and prosodic studies of CS.

Key Results

  • The MCSM corpus analysis showed that Algerian broadcast media had a CS rate of 148.6 segments per hour, compared to 62.5 for Moroccan and 21.2 for Tunisian.
  • The FACST corpus, after initial processing, contains 8 hours of speech, with 4 hours already transcribed. This transcribed data exhibits a high density of code-switching, predominantly short intra-sentential switches, but also longer segments.
  • Experiments with parallel ASR systems indicated that when speech data is processed with the matching language ASR system, higher confidence scores are generally obtained. For instance, the French ASR system showed higher posteriors for French speech segments, and vice-versa for Algerian Arabic.

Limitations and Future Work

  • Lab vs. Field: The authors acknowledge that the FACST corpus collection, while designed to elicit naturalistic CS, was conducted in a soundproof room and is not a direct sociolinguistic survey in a natural field setting.
  • Annotation Underway: The corpus annotation and transcription are still in progress, meaning that the presented statistics are based on a subset of the data.
  • ASR Challenges: Detecting short CS segments remains a significant challenge for current ASR systems.
  • Future Work: The authors plan to continue investigating the use of speech technologies for CS detection and quantification. They aim to further analyze the FACST corpus to gain deeper insights into the acoustic-phonetic and prosodic aspects of French/Algerian Arabic code-switching. They also intend to explore how these technologies can support large-scale CS studies.

Summary updated

BibTeX
@inproceedings{amazouz2017addressing,
  author = {Djegdjiga Amazouz and Martine Adda-Decker and Lori Lamel},
  title = {{Addressing Code-Switching in French/Algerian Arabic Speech}},
  booktitle = {Interspeech},
  year = {2017},
}
Figure