On 28 July 2026, Unisound AI Technology Co., Ltd. (Unisound) released a voluntary announcement detailing a comprehensive upgrade of its speech large language models U2-ASR and U2-TTS. The update strengthens the company’s multimodal capabilities and positions the U2 suite as a global speech-interaction infrastructure for use cases such as overseas expansion, embodied intelligence and agent-to-agent communication.

U2-ASR now supports an additional 13 international languages, broadening coverage to key markets in Europe, Southeast Asia, the Middle East and Latin America. U2-TTS adds multilingual synthesis for eight new Southeast Asian languages. Cumulatively, the U2 platform processes more than 100 Chinese dialects and over 15 international languages through a single set of APIs, lowering development and maintenance barriers for enterprises deploying multi-language speech services.

Performance metrics underscore the upgrade: • U2-ASR records an average Character Error Rate (CER) of 6.58% across 113 languages under unified evaluation standards. • In the ChinaVoices 2026 Chinese Multi-dialect Speech Recognition Challenge, U2-ASR ranked first in the limited-data track with a 9.235% CER and second in the open-data track.

U2-TTS achieved higher intelligibility and naturalness than mainstream industry comparables in subjective testing. A streaming neural-network acoustic model enables chunk-by-chunk real-time output of 24 kHz high-fidelity speech, reducing first-packet latency and supporting time-sensitive interactions such as live voice conversations.

The synchronized upgrade allows U2-ASR and U2-TTS to deliver an end-to-end multilingual speech interaction chain, now fully accessible via Unisound’s TokenHub large language model service platform with standard APIs. The company emphasized its commitment to “AI for Good,” aiming to lower language barriers and broaden access to AI capabilities across global markets.