📄 NatiQ: An End-to-end Text-to-Speech System for Arabic

Summary

The paper presents NatiQ, a web-deployed Arabic text-to-speech system. Arabic is a hard case for TTS because ordinary text leaves out short vowels, so the pronunciation has to be reconstructed before synthesis. The authors recorded two in-house corpora at 44kHz. "Amina" is an expressive female voice reading children's books in Modern Standard Arabic, intended for education. "Hamza" is a neutral male voice reading general and news content, intended for media.

Their pipeline first expands numbers and abbreviations into words and then restores diacritics with Farasa. A language expert reviewed the diacritized text. Transcripts were adjusted to match what the speakers actually said, and the audio was segmented with attention to sentence boundaries.

They compared three acoustic-model/vocoder pairs: Tacotron1+WaveRNN, Tacotron2+WaveGlow and ESPnet Transformer TTS+Parallel WaveGAN. Evaluation combined ASR-based CER/WER, MOS from human listeners and real-time factor. The ESPnet system was fastest and most intelligible overall, while Tacotron1 got the best MOS for the expressive voice. The work shows that near-English TTS quality is reachable for Arabic when careful text preprocessing is in place.

Key points

• Corpora: Amina has 3964 segments and 50,714 words; Hamza has 6005 segments and 80,409 words. Segments average about 7 seconds, or roughly 12 words.
• Test set: 100 sentences from six domains (Culture, Economy, Literature, Politics, Sports, Technology), synthesized by every model in both voices, giving 600 audio files.
• Automatic evaluation: the Arabic ASR of Hussein et al. (2022) transcribed the synthesized audio. ESPnet2 scored best:
  • Hamza: CER 8.01, WER 24.87.
  • Amina: CER 17.47, WER 40.42.
  • Tacotron2 did worst, with WER 93.62 on Hamza.
• MOS: 14 raters (7 female, 7 male) each scored 15 random samples on a 1–5 scale.
  • Best results were ESPnet2 for Hamza (4.40) and Tacotron1 for Amina (4.21).
  • The conclusion instead reports 4.72 and 4.35, which is inconsistent with the tables.
  • Raters preferred the neutral voice over the expressive one.
• Speed (RTF): ESPnet2 reached 0.09 on a V100 GPU, about 1.5× faster than Tacotron2 and about 17× faster than Tacotron1. On CPU it ran at 4.24×RT, so no system was real-time on CPU.

Interesting ideas

• I1. Round-trip evaluation: the generated output is judged by feeding it to an independent recognizer and comparing the result with the input, after stripping vowels so both sides are at the same level of detail.
• I2. Fixing the label instead of the data: when a speaker mispronounced a vowel, the authors edited the transcript to match the audio rather than re-recording, so the pair stayed consistent at low cost.
• I3. Normalization before the main analyzer: digits, fractions, abbreviations and titles are expanded into full words before diacritization, so the downstream tool always receives regular word sequences.
• I4. Targeted expert review: an automatic tool with over 94% accuracy was still checked by a human, focusing on known weak spots like named entities and foreign words.
• I5. Multi-axis evaluation: systems were ranked jointly on intelligibility, human preference and latency, which exposed trade-offs; for example, Tacotron1 was preferred for expressive speech but was the slowest.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (back): text → graph → text output could be reparsed and compared with the source graph, after normalizing away detail the parser cannot recover, such as punctuation or word-order variants.
• I2 (learn): when silver trees and source text disagree on things like typos or nonstandard spellings, the reference annotation could be adjusted to the actual text instead of discarding the sentence.
• I3 (tokens): an explainable pre-pass could expand numerals, units and abbreviations into lexical tokens, with recorded spans, before tagging and lemmatization.
• I4 (learn): LLM-teacher or human review effort could be concentrated on error-prone classes, such as proper names, loanwords and rare verb frames, rather than sampled uniformly.
• I5 (learn): parser and reader variants could be reported on accuracy, explanation quality and throughput together, so a faster model such as gparse is chosen knowingly.