Alibaba launches Qwen-Audio-3.0-Realtime with 4 key upgrades

Alibaba's new real-time voice AI model adds duplex conversation, tool-calling, and emotional awareness — capabilities that could reshape how businesses deploy AI voice agents.

Alibaba Group launched Qwen-Audio-3.0-Realtime, a voice interaction model with duplex conversation and tool-calling, entering a market where its predecessor holds the No. 1 spot on the Speech Arena leaderboard with an Elo of 1,236.

"Voice interaction is moving from command-based to conversational," Wu Jianjun, head of Qwen AI hardware product at Alibaba Group, said in a commentary published Tuesday. "The companies that succeed will build environments where intelligent computing becomes almost invisible."

The model upgrades across four areas: intelligence, Agent tool calling, empathetic conversations, and duplex interaction fluency. Alibaba offers two variants — Plus, with stronger reasoning for complex tasks, and Flash, optimized for low-latency responses. Use cases span intelligent customer service, education and training, entertainment interaction, and emotional companionship. The previous-generation Qwen-Audio-3.0-TTS-Plus model, priced at $27.6 per million characters, edged out SpeechifyAI's Simba 3.2 by two Elo points and Google's Gemini 3.1 Flash TTS by 22 points in blind user evaluations.

The launch strengthens Alibaba's position in the voice AI market against Google, SpeechifyAI, and Cartesia, while deepening the integration between its Qwen foundation models and Alibaba Cloud infrastructure. For investors, the model expands the addressable market for Alibaba's AI services beyond text-based applications into voice-enabled customer service, education, and entertainment — segments where Alibaba Cloud competes with Amazon Web Services and Microsoft Azure for enterprise AI workloads.

The real-time voice model builds on Alibaba's broader AI hardware strategy. The company's Qwen Glasses, announced earlier this year, combine multimodal reasoning from the Qwen model family with Alibaba Cloud's computing infrastructure to enable real-time visual understanding, translation, and contextual assistance. The new voice model could serve as the conversational backbone for such devices, allowing users to interact with AI assistants through natural speech rather than text commands.

Alibaba's Tongyi Lab had already placed its Fun-Realtime-TTS model in the global top five on the Speech Arena leaderboard in May 2026. The Qwen-Audio-3.0-Realtime launch extends that trajectory, adding duplex interaction — where both parties can speak and interrupt naturally — and Agent tool-calling, which allows the model to execute actions such as booking appointments or querying databases on behalf of users.

Voice AI Market Heats Up as Pricing Pressure Mounts

The competitive environment for voice AI is intensifying. Google's Gemini 3.1 Flash TTS ranks third on the leaderboard at 1,214 Elo, while Cartesia's Sonic 3.5 and Inworld's Realtime TTS-2 round out the top five. SpeechifyAI's Simba 3.2, at 1,234 Elo, remains Alibaba's closest rival. Pricing pressure is also mounting: Alibaba's $27.6-per-million-characters rate undercuts many Western competitors, though Google and Amazon have been reducing their own API prices for voice synthesis.

Alibaba shares traded 1.9% higher in Hong Kong on Wednesday following the announcement. The company's cloud intelligence group, which houses its AI models, reported revenue of 31.7 billion yuan ($4.4 billion) in the March quarter, up 13% year-over-year. Voice AI represents a growth vector within that segment: Gartner projects the conversational AI market will reach $18.4 billion by 2028, up from $8.3 billion in 2024. Alibaba trades at roughly 10x forward earnings, a discount to U.S. cloud peers, as investors weigh regulatory headwinds against its AI monetization potential.

This article is for informational purposes only and does not constitute investment advice.