The Voice AI Behind DingTalk Wins the Triple Crown in Listening, Speaking, and Conversing

This week, Alibaba's family of voice large models has achieved remarkable results on Artificial Analysis, a globally authoritative AI benchmarking platform:

Fun-Realtime-ASR ranked first worldwide with an ultra-low word error rate of 1.8%;

Fun-Realtime-AudioChat secured top positions in both speech reasoning and conversational fluency;

Fun-Realtime-TTS-Preview achieved the highest Elo score of 1190 among domestic models.

To date, Alibaba’s voice large models have claimed the top domestic ranking across all three core areas—ASR (speech recognition), TTS (text-to-speech), and Chat (voice conversation)—achieving a historic "grand slam."

Voice intelligence deeply integrated into DingTalk and Wukong user experiences

As key application platforms for Alibaba's large models, DingTalk and Wukong transform advanced voice modeling capabilities into tangible user benefits. Users experience seamless voice intelligence—from online meetings to face-to-face discussions, from text-based workflows to voice interactions—via features such as DingTalk Meetings, AI Note-Taker, IM voice messages, DingTalk A1, Cleer smart headphones, and other hardware devices. These tools effortlessly understand diverse Chinese dialects and industry-specific terminology.

The newly recognized models will be progressively integrated into these products, continuously enhancing user experience and work efficiency.

Three core capabilities of voice interaction: accurate listening, natural speaking, and fluent conversation

Voice large model capabilities can be divided into three layers: accurate listening (ASR, speech-to-text), natural speaking (TTS, text-to-speech), and fluent conversation (Chat, end-to-end speech understanding and dialogue). These three elements work together to deliver a complete voice interaction experience.

Last year, the Fun-ASR large model, jointly developed by DingTalk and Tongyi Lab, was widely deployed across DingTalk products.

Accurate Listening: Precisely capturing complex meetings and professional expressions

In DingTalk Meetings, where multiple participants speak rapidly and use technical jargon, Fun-ASR accurately identifies speech content. After meetings conclude, it automatically generates near-"zero-error" transcripts, eliminating the need for manual note-taking.

In DingTalk A1 series AI hardware and AI Note-Taker, whether recording interviews, lecture notes, or brainstorming sessions, one-click recording produces high-quality, structured text. Combined with speech reasoning capabilities, the AI not only "hears clearly" but also "understands context," automatically summarizing key points, highlighting important content, and generating actionable to-do lists.

One-click voice-to-text conversion for IM messages, supporting all dialects and accents

Within IM messaging, 60-second long voice messages can be instantly converted into highly accurate text, fully supporting regional dialects, accents, and industry-specific terms.

China’s linguistic diversity poses significant challenges for speech recognition. Powered by Alibaba’s voice large models, DingTalk and Wukong now cover eight major dialect groups: Mandarin, Wu, Cantonese, Min, Hakka, Gan, Xiang, and Jin. Their dialect recognition performance across dozens of cities leads the nation, with accuracy rates exceeding 90% for mainstream dialects.

Whether you're a financial professional speaking Shanghainese, a foreign trade entrepreneur using Cantonese, or frontline staff communicating in Sichuanese, Henanese, or Shaanxi dialect, DingTalk can “understand” you.

According to the Artificial Analysis leaderboard, the latest ASR model achieves a word error rate of just 1.8%, further improving transcription accuracy for real-time subtitles and meeting records.

Natural Speaking & Fluent Conversation: Human-like dialogue and intelligent responses

Voice interaction requires not only "accurate listening" but also "natural speaking" and "fluent conversation." On devices like Cleer smart headphones, powerful speech recognition and synthesis capabilities make checking messages, dictating replies, and issuing voice commands during commutes smoother and more natural.

Fun-Realtime-TTS-Preview ranks first among Chinese-developed TTS models with an Elo score of 1190, producing AI-generated speech that sounds more natural and emotionally expressive, significantly enhancing the voice output experience in DingTalk and Wukong products.

Fun-Realtime-AudioChat topped global rankings in both speech reasoning (97.6%) and conversational fluency (97.8%). It understands semantic logic and handles real-world dialogue complexities such as interruptions, follow-up questions, and topic shifts, achieving interaction fluidity close to human levels.

Deeply customized industry language models—truly speaking your professional language

General-purpose models may fail to understand specialized terms like "LPR," "earnout clauses," "Gleevec," or "civil retrial," leading to errors in professional settings.

DingTalk and Alibaba's voice team have collaborated to train models covering vocabulary across mainstream and niche industries—including internet, AI, finance, healthcare, law, automotive, manufacturing, education, and government—and have built deeply customized industry-specific language models.

In finance, the model accurately recognizes terms such as "reverse repo," "MLF," and "asset management新规," significantly improving the quality of meeting minutes and compliance documentation.

In healthcare, complex terms like "methotrexate," "EGFR targeted therapy," and "MDT multidisciplinary consultation" are no longer knowledge blind spots.

In legal practice, concepts ranging from "unjust enrichment" to "bona fide acquisition" are precisely captured in court records and lawyer interviews.

In manufacturing, recognition accuracy for technical terms such as "tolerance fit," "CNC machining," and "PPAP" has greatly improved, doubling productivity in production meetings.

Industry customization goes beyond simply adding words to a dictionary—it involves deep optimization at the language model level, enabling algorithms to truly grasp industry contexts, distinguish homophones, recognize abbreviations, and interpret jargon, ensuring every voice interaction “speaks the lingo and understands the practice.”

Voice is the most natural form of communication. From "hearing accurately" to "understanding meaning," and from "being able to speak" to "engaging in conversation," Alibaba’s voice large model "grand slam" represents not just technological leadership, but a comprehensive leap forward in product experience. DingTalk and Wukong will continue transforming world-leading voice AI capabilities into accessible workplace tools—making meetings more efficient, note-taking effortless, and communication seamless across languages and dialects.

We dedicated to serving clients with professional DingTalk solutions. If you'd like to learn more about DingTalk platform applications, feel free to contact our online customer service or email at This email address is being protected from spambots. You need JavaScript enabled to view it.. With a skilled development and operations team and extensive market experience, we’re ready to deliver expert DingTalk services and solutions tailored to your needs!