Alibaba launches CosyVoice Studio, China's first one-stop AI voice platform

On August 7, Alibaba officially launched CosyVoice Studio, which it calls China's first one-stop AI voice productivity platform. The platform is built on Alibaba's self-developed voice model Qwen-Audio, which ranks first globally in speech recognition (ASR), real-time interaction (RealTime), and speech synthesis (TTS) on the benchmark platform Artificial Analysis.

Voice is becoming the core interface of human-computer interaction in the AI era. As large models mature, users are increasingly choosing voice conversation over text input, and the industry's technical architecture is shifting accordingly: scattered transcription and synthesis tools are converging toward end-to-end voice agents. CosyVoice Studio opens up Alibaba's voice model capabilities to the public for the first time as an aggregated product.

The platform has three core functions. The voice keyboard CosyVoice goes beyond general speech-to-text by layering semantic understanding and text generation on top of transcription: it automatically filters out filler words such as "um" and "uh" and turns scattered thoughts into clear, logical expressions, while self-corrections during speech leave no trace and only the final intent is preserved. In formal work settings, dictating a passage can directly generate emails, work reports, and meeting minutes.

For professional scenarios, CosyVoice supports intelligent transcription of special text such as numbers and formulas, converting spoken "3.58 billion" and "12.6 percent" directly into figures. Financial analysts dictating market briefings, researchers describing experimental data, and journalists organizing interview notes can all get usable structured output. A built-in "notes" mode targets long-form recording scenarios such as meetings, interviews, and classes, offering real-time verbatim transcription, voiceprint-based speaker diarization, automatically generated structured chapters, and one-click multilingual translation, with support for uploading recordings of up to six hours.

CosyAgent is an AI agent builder that lets users quickly create agents with enterprise knowledge, tool calling, and real-time voice interaction through natural language, with deep prompt and workflow configuration for scenarios such as intelligent customer service and telemarketing. The audio content creation tool CosyCreative comes with thousands of voice timbres and supports voice cloning: users upload a document or web link, or type a single sentence, and quickly generate a dialogue podcast or a multi-role audiobook.

Starting today, users can search for "CosyVoice" in app stores on iOS, Android, Mac, and Windows and use it free of charge without limits. CosyAgent and CosyCreative are currently being tested with enterprise users on a whitelist invitation basis and will open up broadly in the near future, according to the official CosyVoice Studio website (cosyvoice.net).

What to watch next: how enterprises adopt CosyAgent and CosyCreative once they fully open, and whether CosyVoice Studio can differentiate in the crowded voice market — especially whether the "semantic understanding plus voice" combination can turn Qwen-Audio's benchmark results into a real product moat.