阿里推出国内首个AI语音平台CosyVoice Studio，将语义理解融入语音能力

TL;DR - Alibaba launched CosyVoice Studio (Aug 7), a one-stop AI speech productivity platform built on its in-house Qwen-Audio model, bundling transcription, audio content generation, and voice agents into a single product. It matters because it packages speech capability as an end-to-end voice-Agent stack rather than isolated ASR/TTS tools, reflecting voice becoming a primary AI interface.

• Claims Qwen-Audio ranked #1 globally on Artificial Analysis across three tracks: ASR, RealTime interaction, and TTS.
• "语音键盘" (voice keyboard) layers semantic understanding and text generation on top of transcription: strips fillers ("嗯", "那个"), resolves self-corrections to final intent, formats emails/meeting notes, and normalizes spoken numbers to symbolic form ("百分之十二点六" → 12.6%).
• "随记" mode does live verbatim transcription with speaker diarization by voiceprint, auto-generated structured chapters, one-click multilingual translation, and up to 6-hour continuous recordings or uploaded files.
• CosyAgent builds voice agents via natural language with enterprise knowledge, tool calling, and prompt/workflow config; CosyCreative offers 1000+ voices plus voice cloning for podcasts/multi-role audiobooks. Apps are free and unlimited on iOS/Android/Mac/Windows; CosyAgent and CosyCreative are whitelist-invite for enterprises.