On September 23, Alibaba published a family of five audio models and lowered its prices by 95% on voice recognition, 70% on synthesis and 85% on real-time interaction.
On September 23, 2026, Alibaba's Qwen team launched the Qwen-Audio 3.1 series, which now includes five models covering sound understanding, generation, interaction and creation. The three historical building blocks — speech recognition, synthesis and real-time dialogue — have been revised, and two newcomers complete them: ASR-Next for the detailed analysis of a recording, TTS-Next for the production of audio content.
ASR-Next can assign each passage to a speaker with a timestamp, and identify emotions, ambient noises or mechanical sounds in the process. TTS-Next, on the other hand, is no longer content to read a text: it simultaneously produces the voice, the effects and the background sound, which brings it closer to a production tool rather than a simple reading engine.
The most visible gesture remains pricing. Alibaba has announced reductions of up to 95% on voice recognition, about 70% on synthesis and nearly 85% on the real-time model. In a market where American laboratories also cut their API prices in September, voice is in turn becoming a cost war ground.
Read the full article on the original source
Go to the source →