Qatar Computing Research Institute (QCRI) has developed NatiQ, an end-to-end text-to-speech (TTS) system for Arabic utilizing encoder-decoder architectures. The system employs Tacotron-based models and Transformer models to generate mel-spectrograms, which are then synthesized into waveforms using vocoders like WaveRNN, WaveGlow, and Parallel WaveGAN. Trained on in-house speech data featuring a neutral male voice (Hamza) and an expressive female voice (Amina), NatiQ achieves a Mean Opinion Score (MOS) of 4.21 and 4.40, respectively. Why it matters: This research advances Arabic language technology, providing high-quality TTS synthesis that can enhance accessibility and usability of digital content for Arabic speakers.
MBZUAI researchers developed LLMVoX, a system enabling LLMs to produce real-time speech, including Arabic. LLMVoX addresses limitations of existing end-to-end and cascaded pipeline approaches, which suffer from either degraded reasoning or latency. LLMVoX was developed as part of Project OMER, which was recently awarded Regional Research Grant from Meta. Why it matters: This enhances the potential of LLMs to function as more natural, multimodal virtual assistants, especially for Arabic-speaking users in the Middle East.
MBZUAI researchers introduce LLMVoX, a 30M-parameter, LLM-agnostic, autoregressive streaming text-to-speech (TTS) system that generates high-quality speech with low latency. The system preserves the capabilities of the base LLM and achieves a lower Word Error Rate compared to speech-enabled LLMs. LLMVoX supports seamless, infinite-length dialogues and generalizes to new languages with dataset adaptation, including Arabic.
ElevenLabs, a voice AI research and product company, presented at MBZUAI's Incubation and Entrepreneurship Center (IEC) on the adoption of audio AI in the Middle East. Hussein Makki, general manager for the Middle East at ElevenLabs, highlighted the potential of voice-native AI across sectors like telecommunications, banking, and education. ElevenLabs focuses on making content accessible and engaging across languages and voices through its text-to-speech models. Why it matters: This signals growing interest and investment in voice AI applications within the region, potentially transforming customer service and content accessibility in Arabic.
Pedro J. Moreno, former head of ASR R&D at Google, presented a talk at MBZUAI on the past, present, and future of speech technologies. The talk covered the evolution of speech tech, his career contributions including work on Google Voice search, and the impact of LLMs on speech science. He also discussed the interplay between foundational and applied research and preparing the next generation of scientists. Why it matters: The talk provides insights into the trajectory of speech technologies from a leading researcher, highlighting future directions and the ethical considerations surrounding AI's impact on society.
The Qatar Computing Research Institute (QCRI) has released QASR, a 2,000-hour transcribed Arabic speech corpus collected from Aljazeera news broadcasts. The dataset features multi-dialect speech sampled at 16kHz, aligned with lightly supervised transcriptions and linguistically motivated segmentation. QCRI also released a 130M word dataset to improve language model training. Why it matters: QASR enables new research in Arabic speech recognition, dialect identification, punctuation restoration, and other NLP tasks for spoken data.
MBZUAI's Hanan Al Darmaki is working to improve automated speech recognition (ASR) for low-resource languages, where labeled data is scarce. She notes that Arabic presents unique challenges due to dialectal variations and a lack of written resources corresponding to spoken dialects. Al Darmaki's research focuses on unsupervised speech recognition to address this gap. Why it matters: Overcoming these challenges can improve virtual assistant effectiveness across diverse languages and enable more inclusive AI applications in the Arabic-speaking world.
A new paper from MBZUAI demonstrates that state-of-the-art speech models can be easily jailbroken using audio perturbations to generate harmful content, achieving success rates of 76-93% on models like Qwen2-Audio and LLaMA-Omni. The researchers adapted projected gradient descent (PGD) to the audio domain to optimize waveforms that push the model towards harmful responses. They propose a defense mechanism based on post-hoc activation patching that hardens models at inference time without retraining. Why it matters: This research highlights a critical vulnerability in speech-based LLMs and offers a practical solution, contributing to the development of more secure and trustworthy AI systems in the region and globally.