📄 Arabic Text-To-Speech (TTS) Data Preparation

Summary

This paper offers practical guidance on building recorded data for Arabic text-to-speech. The authors argue that Arabic lags behind other languages in TTS and that the quality of the human-recorded ground truth largely decides how natural and intelligible the resulting voice will be. They combine a linguistic view with a studio-production view. They first outline a TTS pipeline aimed at banking and finance: text normalization, diacritization, spell checking, utterance segmentation and audio generation. They then describe the data work.

• Selecting the corpus: open-domain or closed-domain text, phonetically balanced, with controlled utterance length.
• Diacritizing the text: partial or full, to remove lexical ambiguity.
• Directing the recording sessions: voice-actor selection, tempo, session length, accent, intonation, pauses and noise.

Finally, they propose ways for listeners to judge speech quality: naturalness, degraded MOS, intelligibility and comprehensibility tests, plus a checklist of perceptual factors. The paper reports no experiments or released dataset. Its value lies in the specifications it gives to voice coaches, actors and annotators working on a low-resource, diacritic-dependent language.

Key points

• The proposed pipeline: normalization turns numbers, dates, abbreviations and foreign words into Arabic text, then diacritization, an Arabic spell checker, utterance segmentation, and finally audio generation. The target domain is finance and banking.
• The data should be phonetically balanced and balanced in length. Phrases are capped at about 12 units, roughly what a speaker can say without pausing.
• The paper tabulates eight Arabic diacritics with IPA and Buckwalter equivalents. Partial diacritization is presented as a cheaper way to resolve ambiguity, for example the many readings of an undiacritized word meaning "study/teach/lesson".
• Recording rules include:
  • at most about 2 hours of recording per day;
  • a 10-minute break every hour;
  • returning to reference sentences after every 50 utterances;
  • re-recording instead of technical cuts;
  • pausing only at punctuation, with short vowels dropped at pauses;
  • keeping the delivery natural rather than acted.
• Evaluation uses four tests: naturalness, degraded MOS with randomized natural and synthetic clips, intelligibility (MOS and WER from listener transcriptions), and comprehensibility (questions scored 0/1).

Interesting ideas

• I1. Morphological analyzers can use whatever diacritics are present as hard filters, discarding any analysis that contradicts them, so sparse annotation still prunes ambiguity cheaply. — For Mova (tokens): Partial cues such as a known lemma, a capitalization hint or a silver POS can act as constraints that eliminate candidate analyses in the tagger, rather than as soft features.
• I2. Speakers periodically go back to a fixed set of reference sentences so that tempo and style do not drift over a long session. — For Mova (learn): A small frozen anchor set can be re-scored at regular intervals during training and self-play to catch drift before it contaminates silver data.
• I3. Defects are fixed by redoing the whole recording, never by splicing fragments, because patches break consistency. — For Mova (back): When generated text from the text → graph → text compressor has a local error, regenerating the whole sentence from the graph should yield more coherent output than patching words in place.
• I4. Intelligibility (writing down what was heard, regardless of meaning) is tested separately from comprehensibility (answering questions about the content). — For Mova (back): The compressor can be scored twice, once on surface reconstruction and once on whether Mova's readers answer the same questions on the regenerated text as on the original.
• I5. Evaluators rate a shuffled mix of real human recordings and synthetic ones, so the human items act as a built-in reference and bias check. — For Mova (learn): Gold human trees and frames can be hidden among model outputs when LLM teachers or witness parsers judge quality, serving as negative and positive controls for the judges themselves.