MediaTek Research, MediaTek's advanced technology R&D unit, today (July 7th) announced the official release of MR Breeze ASR 9, a new open-source speech recognition model. Optimized for common Taiwanese expressions and accents, the model demonstrates even higher recognition accuracy in mixed Chinese and English speech environments. Developed based on OpenAI Whisper and released under the Apache 25 license, this model will accelerate the adoption of AI voice applications by more local developers and industries.
Dr. Da-Shan Hsu, head of MediaTek's Innovation Center, stated that mainstream international speech models previously struggled with understanding Taiwanese speech in the context of Chinese. This included misidentifying "what happened" as "what happened with peanuts," a common error. The MR Breeze ASR 25 was developed to "enable AI to better understand Taiwanese speech."
According to official data, MR Breeze ASR 25 achieves nearly 10% improvement in local speech recognition accuracy compared to OpenAI Whisper, and a significant 56% improvement in code-switching, a common Chinese-English sentence structure in Taiwan. This technology is applicable to common scenarios such as smart customer service, meeting transcript transcription, and subtitle generation, and also opens up new possibilities for future field deployment and device-based AI.
The MR Breeze ASR 25, the second-generation speech recognition model in the MR Breeze series, is part of MediaTek's "Full-Stack Traditional Chinese AI" strategy. The unit has already released a multimodal language model (V-LLM) with visual understanding capabilities, a TTS model specifically tuned for Taiwanese accents, and a language model supporting external function calls, all in 2025, building a comprehensive localized AI technology ecosystem.
Currently MR Breeze ASR 25Now available through HuggingFaceDevelopers and businesses can freely download and use it. Interested developers can visit the MediaTek Innovation Base official platform for more technical details and application cases, or directly try out the model's performance.
As global competition in voice AI technology continues to heat up, MediaTek Innovation Base is demonstrating its long-term commitment and practical capabilities to promoting "Taiwan's language technology independence" by strengthening local language understanding and open source sharing strategies.
