Breeze ASR 26: Bridging the Gap for Taiwanese Hokkien (Taigi) Recognition
MediaTek Research unveils Breeze ASR 26, the first open-source model optimized for Taiwanese Hokkien (Taigi). Part of the MR Breeze 3 series, this 2B parameter model masters code-switching between Mandarin, Taigi, and English, bringing AI closer to Taiwan's unique linguistic reality.
Breeze ASR 26: Bridging the Gap for Taiwanese Hokkien (Taigi) Recognition
In the evolution of speech technology, low-resource languages often get left behind. For Taiwan, while Mandarin recognition has reached maturity, Taiwanese Hokkien (Taigi) has long remained a challenge for global AI models.
Recognizing this gap, MediaTek Research (MR) announced Breeze ASR 26 in early 2026. As the centerpiece of the MR Breeze 3 series, this model marks a historic milestone: the first high-performance, open-source ASR system specifically tuned for the sounds and soul of Taiwan’s mother tongue.
Understanding the Mother Tongue: 2 Billion Parameters
Breeze ASR 26 is a 2-billion parameter model that builds upon the OpenAI Whisper architecture. Unlike its predecessor, Breeze ASR 25 (which focused on Taiwanese Mandarin), ASR 26 was trained on over 10,000 hours of high-quality Taiwanese Hokkien speech data.
Key Technical Specifications:
- Base Architecture: Fine-tuned from OpenAI Whisper
- Model Size: ~2B parameters (~2.9 GB quantized)
- License: Apache 2.0
- Specialization: Taiwanese Hokkien (Taigi) with Trilingual Code-switching support.
Why Breeze ASR 26 is a Game-Changer
What makes Breeze ASR 26 truly remarkable is its ability to handle the complex linguistic patterns of daily life in Taiwan.
1. Mastering "Trilingual" Code-Switching
Taiwanese conversations rarely happen in just one language. It’s common to hear a mix of Mandarin, Taigi, and English in a single breath. Breeze ASR 26 is specifically trained to handle these transitions. Example: "你這個 kha-bang (bag) 有夠媠 (beautiful), 在哪裡買的？" The model can accurately transcribe such "mixed" sentences, a feat that previous models struggled with.
2. Significant Accuracy Breakthrough
Global models like Whisper-large-v2 often exhibit high error rates when transcribing Taigi. Breeze ASR 26 achieves a Character Error Rate (CER) of approximately 30.13% on the Breeze Taigi Benchmark—a massive leap forward compared to general-purpose systems.
3. Optimized for Local Deployment
Privacy is a core value for MediaTek Research. Breeze ASR 26 is optimized to run on consumer hardware. With quantization, it can run on laptops with as little as 4GB of VRAM (like an RTX 3050), allowing for private, offline transcription of sensitive conversations or heritage projects.
4. Standardized Output
To ensure compatibility with modern text processing and LLMs, Breeze ASR 26 transcribes spoken Taigi directly into Traditional Mandarin Chinese characters. This makes the output immediately readable and useful for a wider audience.
The MR Breeze 3 Ecosystem
Breeze ASR 26 is part of a larger family of models released in 2026:
- BreezyVoice 26: A TTS model that speaks with a natural Taiwanese rhythm and intonation (perfect 5.0 MOS score).
- Breeze Guard 26: An AI security model trained on 12,000+ local risk scenarios, including Taiwanese fraud tactics and misinformation.
Conclusion: Preserving Culture Through AI
Breeze ASR 26 is more than just a technical achievement; it is a tool for cultural preservation. By enabling machines to understand and transcribe Taigi with high accuracy, MediaTek Research is ensuring that Taiwan’s linguistic heritage remains relevant in the digital age.
Whether you are building a trilingual customer service bot, transcribing oral histories, or creating accessible subtitles for Taigi content, Breeze ASR 26 is the new standard.
Developers can access the weights on Hugging Face at MediaTek-Research/Breeze-ASR-26.