As the age of AI agents arrives, Unisound continues to evolve its general-purpose intelligent computing foundation, Shanhai Atlas. Following the upgrade of the Shanhai Zhiyi 5.0 medical large model at the end of last year, the company today officially launches Shanhai Zhiyin 2.0, completing another key piece of the capability map for its upgraded "one foundation, two wings" technology strategy.

Shanhai Zhiyin Large Model 2.0: built on the multimodal and cross-lingual foundation capabilities of Shanhai Atlas, it enables specialized vertical AI agents such as Shanhai Zhiyi to benefit households everywhere. Understanding professional terminology and regional accents, bringing warmth and human connection to conversation, and responding with exceptional agility are the three major capability advances in this upgrade.

01 Understanding Expertise and Accents: A Comprehensive ASR Upgrade

Across both public test sets and Unisound's proprietary all-scenario test sets, the model's upgraded ASR capabilities demonstrated leading speech recognition performance. It achieved comprehensive leadership from general to extreme scenarioses, outperforming mainstream open-source and closed-source speech models in China and reaching the highest level in the industry. In particularly challenging conditions involving complex noise and dialectal accents, it improved performance by 2.5% to 3.6% over mainstream ASR models. In complex background-noise environments, recognition accuracy surpassed 90% for the first time in the industry.

Public Test Sets

Proprietary Test Sets

Real-world speech recognition also frequently faces challenges such as unclear recognition of technical terms and broken logic. The biggest highlight of this upgrade is that the model can "understand professional language." By combining context with industry terminology, it can understand every term and instruction in specialized settings, improving recognition accuracy by 30%. "It is not merely hearing words; it is understanding what they mean."

For example, during a test-drive scenario at an automotive 4S dealership, when a salesperson refers to the steering wheel, the model can use logical reasoning to correctly recognize "half-width steering wheel" even if the term has not appeared explicitly in the preceding context.

In high-stakes medical settings, the model can explicitly inject specialized terms such as "epalrestat" and "metformin" for targeted enhancement, ensuring more accurate recognition.

The model also supports recognition and transcription for more than 30 Chinese dialects and 14 international languages. Whether handling challenging Cantonese, Hokkien, and Shanghainese or English, Japanese, Korean, French, German, Thai, and other languages, it delivers accurate transcription. It can further integrate visual semantics from materials such as presentation slides to create a closed-loop audiovisual interaction, improving recognition results even more.

02 Speaking with Warmth and Connection: An Expressive Evolution in TTS

If ASR is the "ears," then TTS is the "voice." Shanhai Zhiyin TTS is built around "highly humanlike expression + creative diversity," giving speech synthesis both realism and creativity while bringing more warmth to technology.

It currently supports 12 dialects (including Cantonese, Sichuanese, and Shanghainese) and 10 foreign languages. It naturally reproduces throat clearing, laughter, and breathing, and can even switch among 12 Mandarin speaking styles, from gentle to polished to friendly. "Technology should not stand aloof; it should speak to you in the way that feels most comfortable."

The model currently supports Cantonese, Sichuanese, Shanghainese, and 12 dialects in total, along with Japanese, Korean, Thai, and 10 foreign languages in total. It can combine generation across dialects, languages, and emotions. Speech rhythm has also been specially optimized for less-resourced languages, including Japanese geminate consonants and Thai tonal variation, producing speech with naturalness approaching that of native speakers.

Additionally, it supports one-sentence voice cloning and podcast-grade long-form synthesis, empowering audio content creation and interactive entertainment.

Large-model-based speech synthesis typically uses Flow Matching to convert speech tokens predicted by a large language model into mel-spectrograms, which are then reconstructed into final audio by a Neural Vocoder. However, this approach generally suffers from high latency. The industry often reduces latency by processing Flow Matching in segments, but the gains are limited and audio quality can be compromised.

To achieve truly high-quality, low-latency streaming speech generation, Unisound developed an innovative Flow Matching module based entirely on causal attention and jointly optimized it with the Neural Vocoder, creating an end-to-end, fully streaming inference architecture. Without sacrificing synthesis quality, this solution significantly reduces system latency: in low-concurrency scenarioses, time to first audio packet has been reduced to under 90 milliseconds, achieving industry-leading real-time interaction performance.

Causal Attention Mechanism

03 Ultra-Responsive Intelligence: End-to-End Full-Duplex Interaction

True intelligent interaction means understanding context, sensing emotion, and responding naturally. The central challenge in achieving smooth full-duplex interaction with an end-to-end model is to perform understanding, decision-making, and generation simultaneously while streaming audio input, while keeping the conversation state coherent at any moment of interruption. Built on its end-to-end interaction engine, Shanhai Zhiyin 2.0 has solved this challenge and elevated full-duplex capabilities to a new level.

It supports interruption at any time, immediate turn-taking, and coherent follow-up questions, creating a conversation as fluid and responsive as chatting with a truly intelligent friend. "This is not question-and-answer. This is conversation."

What Powers It All?

The answer is Unisound's proprietary Shanhai Atlas integrated intelligent computing foundation. It deeply integrates a general multimodal large-model foundation with the Atlas infrastructure, serving as both the foundation for specialized AI agents and the core of the perceptual AI hub. By effectively integrating traditional ASR, TTS, and full-duplex capabilities into an end-to-end large model, it delivers interaction quality and efficiency that conventional cascaded modules cannot achieve.

This time, AI has finally learned how to speak well.