Abstract
NatiQ is an end-to-end text-to-speech (TTS) system specifically optimized for the Arabic language. Our speech synthesizer utilizes an encoder-decoder architecture with attention mechanisms. We evaluated multiple configurations for generating mel-spectrograms from characters, including Tacotron-1, Tacotron-2, and the faster ESPnet Transformer model. These were paired with various vocoders—WaveRNN, WaveGlow, and Parallel WaveGAN—to synthesize final waveforms. To train the models, we utilized two distinct in-house datasets: a neutral male voice ("Hamza") for news and general content, and an expressive female voice ("Amina") for children's literature. Our best-performing systems achieved Mean Opinion Scores (MOS) of 4.21 for Amina and 4.40 for Hamza. Objective evaluations using Word Error Rate (WER) and Real-Time Factor (RTF) indicate that the ESPnet-based end-to-end architecture provides the best balance of accuracy and synthesis speed.
Introduction
High-quality Text-to-Speech (TTS) systems are essential for accessibility, remote education, and media automation. However, developing such systems for Arabic presents unique linguistic challenges. Standard written Arabic is typically undiacritized, meaning short vowels are omitted from the text. This creates significant phonetic ambiguity for machines, as the correct pronunciation must be inferred from context.
NatiQ was developed to address these challenges by providing a robust, end-to-end pipeline that converts raw, undiacritized Arabic text into natural-sounding speech. By integrating advanced linguistic preprocessing with state-of-the-art neural architectures, NatiQ bridges the gap between written text and high-fidelity audio, particularly for educational contexts accelerated by the COVID-19 pandemic.
Methodology and System Architecture
The NatiQ system follows a two-stage synthesis process: first, converting input text into a mel-spectrogram, and second, converting that spectrogram into a time-domain waveform (vocoding).
Model Configurations
We experimented with three primary architectural pairings to identify the optimal balance of quality and performance:
- Tacotron1 + WaveRNN: A recurrent neural network (RNN) based sequence-to-sequence model.
- Tacotron2 + WaveGlow: An advanced RNN model utilizing location-sensitive attention and a flow-based generative vocoder.
- ESPnet Transformer + Parallel WaveGAN: A transformer-based architecture that utilizes self-attention mechanisms. This configuration allows for parallelization during training and faster inference compared to sequential RNN models.
Linguistic Preprocessing
To resolve the ambiguity of undiacritized text, we integrated Farasa, a leading Arabic linguistic tool, to automatically predict and insert diacritics. Furthermore, we implemented a custom text normalization layer to handle non-textual elements. This layer ensures that dates, abbreviations, and numerical values (e.g., "16.43") are correctly expanded into their full phonetic Arabic equivalents before being passed to the encoder.
Datasets
The models were trained on two high-quality, in-house corpora recorded at 44kHz and segmented into approximately 7-second clips:
- Hamza (Neutral): A male speaker providing clear, monotonic narration suitable for news and formal broadcasts. This dataset contains approximately 6,000 segments (80,000 words).
- Amina (Expressive): A female speaker providing emotive narration for children's storybooks. This dataset contains approximately 4,000 segments (50,000 words).
To ensure high data quality, the authors employed a "transcript matching" strategy: if a speaker deviated slightly from the script during recording, the text transcript was updated to match the audio, preserving the natural flow of the speech while maintaining alignment for the model.
Results and Discussion
Subjective Evaluation
We conducted Mean Opinion Score (MOS) tests where human listeners rated the naturalness of the synthesized audio on a scale of 1 to 5. The neutral male voice (Hamza) achieved a peak MOS of 4.40 using the ESPnet2 configuration. The expressive female voice (Amina) achieved a peak MOS of 4.21 using Tacotron1. Scores above 4.0 indicate that the synthesized speech is nearing human-level naturalness.
Objective Evaluation and Efficiency
Objective performance was measured using Word Error Rate (WER) via an Automatic Speech Recognition (ASR) system and the Real-Time Factor (RTF) on a GPU.
- Accuracy: ESPnet2 demonstrated the highest accuracy, yielding a WER of 24.87% for the neutral voice, significantly outperforming the Tacotron variants.
- Speed: The ESPnet Transformer model was the most efficient, achieving an RTF of 0.09. This indicates the system can generate 10 seconds of audio in less than one second, making it roughly 17 times faster than the Tacotron1 implementation.
The Challenge of Expressivity
A key finding was the performance gap between neutral and expressive speech. The neutral voice consistently outperformed the expressive voice in objective metrics. Expressive speech involves complex, non-monotonic prosody and emotional nuances that are inherently more difficult for neural networks to model accurately compared to the stable patterns of news narration.
Conclusion
NatiQ demonstrates that end-to-end transformer-based architectures, when combined with sophisticated Arabic diacritization tools, can produce high-quality, efficient Arabic speech synthesis. While the system excels at neutral narration, future research will focus on better disentangling style and content. By modeling emotional tone as a separate objective function, we aim to improve the phonetic clarity of expressive speech without sacrificing its emotive qualities.
Summary updated