What Ant Group open-sourced this time isn’t just a single model—it’s an entire… · @tonychuhai · TwKit

@tonychuhai

What Ant Group open-sourced this time isn’t just a single model—it’s an entire speech production pipeline. AntLingAGI rolled out 5 models under Ming-3.0 in one go, spanning three tracks: TTS for speaking, Audio for comprehension and dialogue, and ASR for transcription. These three tracks can be used independently or chained together end-to-end. The most compelling data comes from third parties. On Artificial Analysis’s open-source TTS leaderboard, when evaluated and compared across 8 standardized cloned voices, Ming-3.0-tts scored an Elo of 1069, taking the #1 spot and landing in the same tier as runner-up Fish Audio Drama 3 Preview (1059). Its English pronunciation accuracy reached 93.2%, leading the chart ahead of ElevenLabs v4 (91.7%). The architectural philosophy behind the model sizing is equally clear: The smallest TTS model is only 0.5B. The long-audio version has 7.9B total parameters with only 1.3B activated. The comprehension model has 126B total parameters with only 6.5B activated. Using MoE (Mixture of Experts) to activate only a small fraction of parameters drastically cuts inference costs—the goal is large-scale production deployment, not just benchmark flexing. It doesn’t sweep every category, either. According to official charts, Audio-flash scored 86.74 on VoiceBench, falling short of Gemini 3.8 Flash’s 92.30; its ASR dialect error rate was 8.42%, slightly higher than Fun-ASR 1.5’s 8.00%. My take: Now that open-source TTS can go toe-to-toe with top commercial voices, monetizing purely on "pleasant synthetic voices" will become increasingly difficult. What will retain paying customers are licensed voice IP libraries, end-to-end long-form production workflows, and distribution channels. If mainstream TTS pricing doesn't soften within the next six months, it simply means I overestimated the impact of open source on the paid commercial market. https://x.com/AntLingAGI/status/2112210816149897641/video/1

曝光 306 · 评论 0 · 点赞 4 · 书签 1