QASR: QCRI Aljazeera Speech Resource -- A Large Scale Annotated Arabic Speech Corpus

Hamdy Mubarak, Amir Hussein, Shammur Absar Chowdhury, Ahmed Ali

TL;DR

The authors introduce QASR (QCRI Aljazeera Speech Resource), a massive 2,000-hour multi-dialect Arabic speech corpus featuring lightly supervised transcriptions and rich metadata. Unlike previous datasets, QASR provides linguistically motivated segmentation, punctuation, and speaker information, effectively bridging the gap between Automatic Speech Recognition (ASR) and downstream Natural Language Processing (NLP) tasks.

What Problem Does This Paper Address?

While Automatic Speech Recognition (ASR) has advanced significantly, speech datasets often strip away or ignore elements crucial for Natural Language Processing (NLP). Spoken language contains unique challenges—such as disfluencies, hesitations, overlapping speech, and code-switching—that are rarely present in clean, well-punctuated text corpora.

For Arabic, this problem is compounded by diglossia: the co-existence of Modern Standard Arabic (MSA) and various regional Dialectal Arabic (DA) variants. Existing Arabic speech datasets typically lack punctuation, speaker metadata, or logical linguistic boundaries. The authors address this disconnect by releasing a large-scale dataset that preserves the link between acoustic speech and structural, semantic text.

Methods

The creation of QASR involved a multi-step pipeline to transform raw broadcast media into a structured machine-learning resource:

  1. Data Collection & Parsing: The authors collected 11 years of archives (2004–2015) from the Aljazeera news channel, encompassing over 4,000 episodes across 19 different programs. They automatically parsed highly inconsistent metadata (e.g., standardizing 9 different spelling variations of "Barack Obama").
  2. Speech-to-Text Alignment: To map the manual text transcripts to the audio timeline, the authors ran an ASR system over the episodes and aligned the output with the text using a modified Smith-Waterman algorithm. This modification allowed for "approximate matches" based on Levenshtein distance to accommodate human transcription errors and repetitions.
  3. Linguistically Motivated Segmentation: Instead of cutting audio arbitrarily based on silence or fixed time limits, the authors combined acoustic cues (silence > 150ms) with strict linguistic rules. Using the Arabic NLP tool Farasa, they created a "NO-STOP-LIST" of 2,200 words to ensure segments never split in the middle of Named Entities, Noun Phrases, or Adjective Phrases.

The authors also established baseline models for five downstream tasks: ASR (using an End-to-End Transformer), Automatic Punctuation Restoration (AraBERT + BiLSTM), Speaker Verification, Arabic Dialect Identification, and Named Entity Recognition (NER).

Datasets

The paper introduces and utilizes the following datasets:

  • QASR Corpus: 2,041 hours of speech (sampled at 16kHz) containing 1.6 million segments, 14.3 million words, and 27,977 unique speakers.
  • Language Model Text: A supplementary dataset of 130 million words to aid in training better language models.
  • Updated MGB-2 Test Set: The authors reused the popular MGB-2 test set but manually enriched it with speaker country, gender, and NER annotations to create a robust evaluation benchmark.

Key Insights and Ideas

  • Segmentation Geometry Matters: Traditional speech datasets like MGB-2 rely heavily on silence for segmentation, resulting in a highly right-skewed segment duration distribution. QASR’s linguistically motivated approach yields a clean bell curve that closely mimics expert human transcription distributions.
Duration distributions of the speech segments for MGB-2, QASR, and test dataset
  • The Comma Dilemma in Arabic: Punctuation restoration models performed significantly worse on commas than on full stops or question marks. This reflects a broader linguistic reality: the use of commas in written Arabic is highly fluid and frequently substituted by full stops or coordinating conjunctions.
  • Speech Transcripts Degrade NLP Performance: Standard NLP tools trained on clean text struggle when deployed on speech transcripts. The authors observed a sharp drop in Named Entity Recognition performance, primarily due to alignment errors and spoken disfluencies (e.g., "OK you are isn't it I mean are you not ashamed?").

Key Results

  • ASR Performance: The baseline End-to-End Transformer trained on QASR achieved a competitive Word Error Rate (WER) of 14.7% on the test set, closely trailing a model trained on the highly optimized MGB-2 corpus (14.3%).
  • Punctuation Restoration: The AraBERT-BiLSTM baseline achieved an F1-score of 59.7% for full stops (.), 54.9% for question marks (?), and 46.7% for commas (,).
  • Speaker Verification: The system achieved an Equal Error Rate (EER) of 7.5% for guest speakers and 9.2% for anchors. Differentiating anchors proved harder because they share identical acoustic and recording environments.
  • NER Drop: Testing the Farasa NER tool on standard text (ANERcorp) yielded an 84.3% F1-score, but this plummeted to 69.8% when evaluated on the QASR speech transcriptions.

Limitations and Future Work

A key limitation identified in the baseline ASR model was a high rate of deletion errors for audio segments longer than 7 seconds. Because QASR's linguistic segmentation naturally favors shorter, uniform chunks, the model struggled to generalize to the longer segments present in the test set.

For future work, the authors suggest exploring multimodal architectures for punctuation restoration that combine lexical features with acoustic cues (like pitch and pause duration). They also plan to focus on enhancing NER models specifically tailored to handle the messy realities of speech transcripts.

Summary updated

Abstract

We introduce the largest transcribed Arabic speech corpus, QASR, collected from the broadcast domain. This multi-dialect speech dataset contains 2,000 hours of speech sampled at 16kHz crawled from Aljazeera news channel. The dataset is released with lightly supervised transcriptions, aligned with the audio segments. Unlike previous datasets, QASR contains linguistically motivated segmentation, punctuation, speaker information among others. QASR is suitable for training and evaluating speech recognition systems, acoustics- and/or linguistics- based Arabic dialect identification, punctuation restoration, speaker identification, speaker linking, and potentially other NLP modules for spoken data. In addition to QASR transcription, we release a dataset of 130M words to aid in designing and training a better language model. We show that end-to-end automatic speech recognition trained on QASR reports a competitive word error rate compared to the previous MGB-2 corpus. We report baseline results for downstream natural language processing tasks such as named entity recognition using speech transcript. We also report the first baseline for Arabic punctuation restoration. We make the corpus available for the research community.
BibTeX
@inproceedings{mubarak2021qasr,
  author = {Hamdy Mubarak and Amir Hussein and Shammur Absar Chowdhury and Ahmed Ali},
  title = {{QASR: QCRI Aljazeera Speech Resource -- A Large Scale Annotated Arabic Speech Corpus}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2021},
}

Loading related content...

Figure