MediaTek Research today (March 4th) unveiled the new MediaTek Research Breeze 3 (MR Breeze 3) series models. This upgrade not only continues the Breeze series' original intention of focusing on Traditional Chinese and Taiwanese culture, but also extends the reach of AI into the field of Taiwanese speech for the first time. At the same time, it creates a localized AI content security defense line for Taiwan's unique social context.
Among them, the Taiwanese speech recognition model Breeze ASR 26 and the security protection model Breeze Guard 26 will be shared with the public in an open-source format. All models in the series are currently available online.Breeze AI LINE Official AccountFor the public to experience.
Breeze ASR 26: Overcoming the challenges of tones and mixed usage, enabling AI to truly understand Taiwanese.
Taiwanese has rich tonal variations and diverse writing systems, which has always been a major challenge for speech recognition technology. While mainstream speech recognition models around the world are becoming increasingly proficient in recognizing English and Mandarin, they often find themselves unable to recognize Taiwanese.
To address this pain point, MediaTek Innovation Base launched Breeze ASR 26. This model is based on the OpenAI Whisper architecture and was trained using approximately 10000 hours of Taiwanese synthesized speech data. Compared to traditional, rigid reading-out data, this batch of synthesized speech covers conversational scenarios that are closer to daily life, with natural speech rate variations and colloquial vocabulary, and even includes the "mixed use of Mandarin and Taiwanese" that Taiwanese people are very accustomed to.
Based on these characteristics, Breeze ASR 26 is able to truly learn the spoken Taiwanese habits of everyday speakers. In practical applications, even when Mandarin or English words are mixed in during speech (e.g., "Your kha-bang is so good, where did you buy it?"), the model can still immediately and accurately recognize the speech and convert it into text.
BreezyVoice 26: Received a perfect score; the AI spoke with a genuine Taiwanese accent.
In addition to being able to understand, it also needs to speak naturally. Following the highly acclaimed BreezyVoice, which synthesized natural Taiwanese Mandarin last year, the newly released BreezyVoice 26 takes this a step further by enabling AI to "speak Taiwanese".
BreezyVoice 26 is developed based on the CosyVoice 2 architecture and is also trained using a large amount of Taiwanese synthesized speech data, which can convert text into natural and fluent Taiwanese speech.
The most notable feature of this model is its stunningly natural speech quality. In the human-evaluated Text-to-Speech Quality (MOS) test, BreezyVoice 26 received a perfect score of 5. Reviewers unanimously agreed that its translated speech sounded just like a real person speaking, mainly thanks to the model's faithful mastery of the intonation and rhythm of Taiwanese speakers from a large amount of synthetic data.
Breeze Guard 26: A cybersecurity gatekeeper specifically designed for the Taiwanese social context.
Beyond language proficiency, the security of AI applications is equally crucial. Most existing international security models are designed for English content, and often struggle to accurately assess risk scenarios unique to Taiwan (such as common local scams and discriminatory remarks within the context of local culture).
To this end, MediaTek Innovation Hub developed the Breeze Guard 26 AI security protection model. Since MR Breeze 2 had already extensively studied Traditional Chinese news, forums, and public documents during its pre-training phase, it possesses a deep understanding of Taiwanese society and culture. Breeze Guard 26 is based on the Breeze 2 8B model, further incorporating over 12000 training data points specifically designed for risk scenarios in Taiwan for specialized training.
This protective network can accurately target common harmful content in Taiwan, including different types of textual content such as vulgar language, fraud, misinformation, verbal discrimination, and political manipulation.
Promoting the open source ecosystem and the LINE official account experience
To continuously build comprehensive local AI capabilities in Taiwan and promote technological development, MediaTek Innovation Base has also released the "Breeze Taigi Benchmark" for Taiwanese speech evaluation and the "TS-Bench" for Taiwan security evaluation, for use by academic and industry partners.
For general users, all three models are currently integrated into [the platform/system].Breeze AI LINE Official AccountUsers can easily switch between three main functions via the menu: "Listen to Me" (Taiwanese speech to text), "Read to Me" (text to Taiwanese speech), and "Content Control" (assessing content security).
