File size: 17,056 Bytes
148368a c3ec9ca 4c29f5e c3ec9ca 0934565 c3ec9ca 3fb7e9e c3ec9ca 3fb7e9e c3ec9ca 5756d9d 13c405b 02de5b7 13c405b 02de5b7 13c405b 0469dbf 13c405b 0469dbf 13c405b 0469dbf 18d0f01 13c405b 18d0f01 3fb7e9e 02de5b7 18d0f01 c3ec9ca 5756d9d c3ec9ca 5756d9d c3ec9ca 5756d9d c3ec9ca 5756d9d c3ec9ca 5756d9d c3ec9ca 18d0f01 c3ec9ca 5756d9d c3ec9ca 18d0f01 9847294 609e630 9847294 c3ec9ca 5756d9d c3ec9ca 6ac52e9 c3ec9ca 3fb7e9e 0469dbf e3aa6c7 0469dbf 5756d9d e3aa6c7 0469dbf 5756d9d 0469dbf 5756d9d 0469dbf fd168fe 0469dbf fd168fe 0469dbf fd168fe 0469dbf 5756d9d 0469dbf fd168fe 0469dbf fd168fe 0469dbf fd168fe 0469dbf 3fb7e9e 983b407 3fb7e9e 5756d9d 18d0f01 c3ec9ca 18d0f01 668602f 18d0f01 c3ec9ca 148368a c3ec9ca |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 |
<!doctype html>
<html>
<head>
<meta charset="utf-8" />
<meta name="viewport" content="width=device-width" />
<title>Iqra’Eval Shared Task</title>
<link rel="stylesheet" href="style.css" />
</head>
<body>
<div class="card">
<h1>Iqra’Eval Shared Task</h1>
<div style="text-align:center; margin: 20px 0;">
<img src="IqraEval.png" alt="" style="max-width:100%; height:auto;" />
</div>
<!-- Overview Section -->
<h2>Overview</h2>
<p>
<strong>Iqra'Eval</strong> is a shared task aimed at advancing <strong>automatic assessment of Qur’anic recitation pronunciation</strong> by leveraging computational methods to detect and diagnose pronunciation errors. The focus on Qur’anic recitation provides a standardized and well-defined context for evaluating Modern Standard Arabic (MSA) pronunciation.
</p>
<p>
Participants will develop systems capable of Detecting Mispronunciations (e.g., substitution, deletion, or insertion of phonemes).
</p>
<!-- Timeline Section -->
<h2>Timeline</h2>
<ul>
<li><strong>June 1, 2025</strong>: Official announcement of the shared task</li>
<li><strong>June 10, 2025</strong>: Release of training data, development set (QuranMB), phonetizer script, and baseline systems</li>
<li><strong>July 24, 2025</strong>: Registration deadline and release of test data</li>
<li><strong>July 27, 2025</strong>: End of evaluation cycle (test set submission closes)</li>
<li><strong>July 30, 2025</strong>: Final results released</li>
<li><strong>August 15, 2025</strong>: System description paper submissions due</li>
<li><strong>August 22, 2025</strong>: Notification of acceptance</li>
<li><strong>September 5, 2025</strong>: Camera-ready versions due</li>
</ul>
<!-- Task Description -->
<h2>Task Description: Quranic Mispronunciation Detection System</h2>
<p>
The aim is to design a model to detect and provide detailed feedback on mispronunciations in Quranic recitations.
Users read aloud vowelized Quranic verses; This model predicts the phoneme sequence uttered by the speaker, which may contain mispronunciations.
Models are evaluated on the <strong>QuranMB.v2</strong> dataset, which contains human‐annotated mispronunciations.
</p>
<div class="centered">
<img src="task.png" alt="System Overview" style="max-width:100%; height:auto;" />
<p><em>Figure: Overview of the Mispronunciation Detection Workflow</em></p>
</div>
<h3>1. Read the Verse</h3>
<p>
The user is shown a <strong>Reference Verse</strong> (What should have been said) in Arabic script along with its corresponding <strong>Reference Phoneme Sequence</strong>.
</p>
<p><strong>Example:</strong></p>
<ul>
<li><strong>Arabic:</strong> إِنَّ الصَّفَا وَالْمَرْوَةَ مِنْ شَعَائِرِ اللَّهِ</li>
<li>
<strong>Phoneme:</strong>
<code>< i n n a SS A f aa w a l m a r w a t a m i n $ a E a a < i r i l l a h i</code>
</li>
</ul>
<h3>2. Save Recording</h3>
<p>
The user recites the verse aloud; the system captures and stores the audio waveform for subsequent analysis.
</p>
<h3>3. Mispronunciation Detection</h3>
<p>
The stored audio is fed into a <strong>Mispronunciation Detection Model</strong>.
This model predicts the phoneme sequence uttered by the speaker, which may contain mispronunciations.
</p>
<p><strong>Example of Mispronunciation:</strong></p>
<ul>
<li><strong>Reference Phoneme Sequence (What should have been said):</strong> <code>< i n n a SS A f aa w a l m a r w a t a m i n $ a E a a < i r i l l a h i</code></li>
<li><strong>Model Phoneme Prediction (What is predicted):</strong> <code>< i n n a SS A f aa w a l m a r w a t a m i n s a E a a < i r u l l a h i</code></li>
<li>
<strong>Annotated Phoneme Sequence (What is said):</strong>
<code>< i n n a SS A f aa w a l m a r w a m i n <span class="highlight">s</span> a E a a < i <span class="highlight">r u</span> l l a h i</code>
</li>
</ul>
<p>
In this case, the phoneme <code>$</code> was mispronounced as <code>s</code>, and <code>i</code> was mispronounced as <code>u</code>.
</p>
<p>
The annotated phoneme sequence indicates that the phoneme <code>ta</code> was omitted, but the model failed to detect it.
</p>
<h2>Training Dataset: Description</h2>
<p>
All data are hosted on Hugging Face. Two main splits are provided:
</p>
<ul>
<li>
<strong>Training set:</strong> 79 hours of Modern Standard Arabic (MSA) speech, augmented with multiple Qur’anic recitations.
<br />
<code>df = load_dataset("IqraEval/Iqra_train", split="train")</code>
</li>
<li>
<strong>Development set:</strong> 3.4 hours reserved for tuning and validation.
<br />
<code>df = load_dataset("IqraEval/Iqra_train", split="dev")</code>
</li>
</ul>
<p>
<strong>Column Definitions:</strong>
</p>
<ul>
<li><code>audio</code>: Speech Array.</li>
<li><code>sentence</code>: Original sentence text (may be partially diacritized or non-diacritized).</li>
<li><code>index</code>: If from the Quran, the verse index (0–6265, including Basmalah); otherwise <code>-1</code>.</li>
<li><code>tashkeel_sentence</code>: Fully diacritized sentence (auto-generated via a diacritization tool).</li>
<li><code>phoneme</code>: Phoneme sequence corresponding to the diacritized sentence (Nawar Halabi phonetizer).</li>
</ul>
<p>
<strong>Data Splits:</strong>
<br />
• Training (train): 79 hours total<br />
• Development (dev): 3.4 hours total
</p>
<!-- Additional TTS Data -->
<h2>Training Dataset: TTS Data (Optional Use)</h2>
<p>
We also provide a high-quality TTS corpus for auxiliary experiments (e.g., data augmentation, synthetic pronunciation error simulation). This TTS set can be loaded via:
</p>
<ul>
<li><code>df_tts = load_dataset("IqraEval/Iqra_TTS")</code></li>
</ul>
<h2>Test Dataset: QuranMB_v2</h2>
<p>
To construct a reliable test set, we select 98 verses from the Qur’an, which are read aloud by 18 native Arabic speakers (14 females, 4 males), resulting in approximately 2 hours of recorded speech. The speakers were instructed to read the text in MSA at their normal tempo, disregarding Qur’anic tajweed rules, while deliberately producing the specified pronunciation errors. To ensure consistency in error production, we developed a custom recording tool that highlighted the modified text and displayed additional instructions specifying the type of error. Before recording, speakers were required to silently read each sentence to familiarize themselves with the intended errors before reading them aloud. After recording, three linguistic annotators verified and corrected the phoneme sequence and flagged all pronunciation errors for evaluation.
</p>
<ul>
<li><code>df_test = load_dataset("IqraEval/Iqra_QuranMB_v2")</code></li>
</ul>
<!-- Resources & Links -->
<h2>Resources</h2>
<ul>
<li>
<a href="https://huggingface.co/datasets/IqraEval/Iqra_train" target="_blank">
Training & Development Data on Hugging Face
</a>
</li>
<li>
<a href="https://huggingface.co/datasets/IqraEval/Iqra_TTS" target="_blank">
IqraEval TTS Data on Hugging Face
</a>
</li>
<li>
<a href="https://github.com/Iqra-Eval/interspeech_IqraEval" target="_blank">
Baseline systems & training scripts (GitHub)
</a>
</li>
</ul>
<p>
<em>
For detailed instructions on data access, phonetizer installation, and baseline usage, please refer to the <a href="https://github.com/Iqra-Eval" target="_blank">
GitHub
</a>.
</em>
</p>
<!-- Submission Details -->
<h2>Submission Details (Draft)</h2>
<p>
Participants are required to submit a CSV file named <code>submission.csv</code> containing the predicted phoneme sequences for each audio sample. The file must have exactly two columns:
</p>
<ul>
<li><strong>ID:</strong> Unique identifier of the audio sample.</li>
<li><strong>Labels:</strong> The predicted phoneme sequence, with each phoneme separated by a single space.</li>
</ul>
<p>
Below is a minimal example illustrating the required format:
</p>
<pre>
ID,Labels
0000_0001, i n n a m a a y a k h a l l a h a m i n ʕ i b a a d i h u l ʕ u l a m
0000_0002, m a a n a n s a k h u m i n i ʕ a a y a t i n
0000_0003, y u k h i k u m u n n u ʔ a u ʔ a m a n a t a n m m i n h u
…
</pre>
<p>
The first column (ID) should match exactly the audio filenames (without extension). The second column (Labels) is the predicted phoneme string.
</p>
<p>
<strong>Important:</strong>
<ul>
<li>Use UTF-8 encoding.</li>
<li>Do not include extra spaces at the start or end of each line.</li>
<li>Submit a single CSV file (no archives). Filename must be <code>teamID_submission.csv</code>.</li>
</ul>
</p>
<h2>Evaluation Criteria</h2>
<p>
IqraEval Leaderboard rankings will primarily be based on the <strong>phoneme-level F1-score</strong>.
</p>
<p>
In addition, we adopt a hierarchical evaluation structure, <a href="https://arxiv.org/pdf/2310.13974" target="_blank">MDD Overview</a>, that breaks down performance into detection and diagnostic phases.
</p>
<p>
<strong>Hierarchical Evaluation Structure:</strong>
The hierarchical mispronunciation detection process relies on three sequences:
<ul>
<li><em>What is said</em> (the <strong>annotated phoneme sequence</strong> from human annotation),</li>
<li><em>What is predicted</em> (the <strong>model’s phoneme output</strong>),</li>
<li><em>What should have been said</em> (the <strong>reference phoneme sequence</strong>).</li>
</ul>
By comparing these three sequences, we compute the following counts:
</p>
<ul>
<li><strong>True Acceptance (TA):</strong>
Number of phonemes that are annotated as correct and also recognized as correct by the model.
</li>
<li><strong>True Rejection (TR):</strong>
Number of phonemes that are annotated as mispronunciations and correctly predicted as mispronunciations.
(These labels are further used to measure diagnostic errors by comparing the prediction to the canonical reference.)
</li>
<li><strong>False Rejection (FR):</strong>
Number of phonemes that are annotated as correct but wrongly predicted as mispronunciations.
</li>
<li><strong>False Acceptance (FA):</strong>
Number of phonemes that are annotated as mispronunciations but misclassified as correct pronunciations.
</li>
</ul>
<p>
From these counts, we derive three rates:
<ul>
<li><strong>False Rejection Rate (FRR):</strong>
FRR = FR/(TA + FR)
(Proportion of correctly pronounced phonemes that were mistakenly flagged as errors.)
</li>
<li><strong>False Acceptance Rate (FAR):</strong>
FAR = FA/(FA + TR)
(Proportion of mispronounced phonemes that were mistakenly classified as correct.)
</li>
<li><strong>Diagnostic Error Rate (DER):</strong>
DER = DE/(CD + DE)
where DE is the number of misdiagnosed phonemes and CD is the number of correctly diagnosed ones.
</li>
</ul>
</p>
<p>
In addition to these hierarchical measures, we compute the standard <strong>Precision</strong>, <strong>Recall</strong>, and <strong>F-measure</strong> for mispronunciation detection:
<ul>
<li><strong>Precision:</strong>
Precision = TR/(TR + FR)
(Of all phonemes predicted as mispronounced, how many were actually mispronounced?)
</li>
<li><strong>Recall:</strong>
Recall = TR/(TR + FA)
(Of all truly mispronounced phonemes, how many did we correctly detect?)
</li>
<li><strong>F1-score:</strong>
F1-score = 2 * Precision * Recall / (Precision + Recall)
</li>
</ul>
</p>
<h2>Potential Research Directions</h2>
<ol>
<li>
<strong>Advanced Mispronunciation Detection Models</strong><br>
Apply state-of-the-art self-supervised models (e.g., Wav2Vec2.0, HuBERT), using variants that are pre-trained/fine-tuned on Arabic speech. These models can then be fine-tuned on Quranic recitations to improve phoneme-level accuracy.
</li>
<li>
<strong>Data Augmentation Strategies</strong><br>
Create synthetic mispronunciation examples using pipelines like
<a href="https://arxiv.org/abs/2211.00923" target="_blank">SpeechBlender</a>.
Augmenting limited Arabic/Quranic speech data helps mitigate data scarcity and improves model robustness.
</li>
<li>
<strong>Analysis of Common Mispronunciation Patterns</strong><br>
Perform statistical analysis on the QuranMB dataset to identify prevalent errors (e.g., substituting similar phonemes, swapping vowels).
These insights can drive targeted training and tailored feedback rules.
</li>
</ol>
<!-- Placeholder for Future Details -->
<h2>Future Updates</h2>
<p>
Further details on <strong>evaluation criteria</strong> (exact scoring weights), <strong>submission templates</strong>, and any clarifications will be posted on the shared task website when test data are released (June 5, 2025). Stay tuned!
</p>
<h2>References</h2>
<ul>
<li>
El Kheir, Y., et al.
"<a href="https://arxiv.org/abs/2211.00923" target="_blank">SpeechBlender: Speech Augmentation Framework for Mispronunciation Data Generation</a>,"
<em>arXiv preprint arXiv:2211.00923</em>, 2022.
</li>
<li>
Al Harere, A., & Al Jallad, K.
"<a href="https://arxiv.org/abs/2305.06429" target="_blank">Mispronunciation Detection of Basic Quranic Recitation Rules using Deep Learning</a>,"
<em>arXiv preprint arXiv:2305.06429</em>, 2023.
</li>
<li>
Aly, S. A., et al.
"<a href="https://arxiv.org/abs/2111.01136" target="_blank">ASMDD: Arabic Speech Mispronunciation Detection Dataset</a>,"
<em>arXiv preprint arXiv:2111.01136</em>, 2021.
</li>
<li>
Moustafa, A., & Aly, S. A.
"<a href="https://arxiv.org/abs/2111.06331" target="_blank">Towards an Efficient Voice Identification Using Wav2Vec2.0 and HuBERT Based on the Quran Reciters Dataset</a>,"
<em>arXiv preprint arXiv:2111.06331</em>, 2021.
</li>
<li>
El Kheir, Y., et al.
"<a href="https://arxiv.org/pdf/2310.13974" target="_blank">Automatic Pronunciation Assessment - A Review</a>,"
<em>arXiv preprint arXiv:2310.13974</em>, 2021.
</li>
</ul>
</div>
</body>
</html>
|