Ant Group Open-Sources Ming-3.0-tts: 93.2% Pronunciation Accuracy Leading Global Open Speech Models

Background and the Problem

With the proliferation of real-time voice agents and automated audio storytelling, demand for low-latency, naturalistic Text-to-Speech (TTS) models has surged. However, high-quality voice generation has historically remained trapped behind proprietary commercial APIs (such as ElevenLabs), imposing steep per-character tariffs and restrictive rate limits. Meanwhile, existing open-source speech engines frequently struggle with robotic cadence, phoneme mispronunciations on domain-specific terminology, and prosody collapse over long paragraphs. Ant Group's newly released Ming-3.0-tts changes the competitive landscape with top-tier benchmark scores.

Architecture and How It Works

1. Dual-Tier Architecture (Dense 0.5B & MoE 7.9B):

• Includes a lightweight 0.5B dense model tailored for edge execution and sub-50ms streaming latency on commodity hardware;
• Features a 7.9B Mixture-of-Experts (MoE) flagship model engineered for nuanced emotional inflections, accent transitions, and realistic human respiration pauses;

1. Zero-Shot Voice Cloning & Affective Control:

• Replicates speaker acoustics from a brief 3–5 second audio reference while allowing fine-grained modulation of speech sentiment and delivery style;

1. Permissive MIT Licensing:

• Released fully under the MIT license on Hugging Face, enabling frictionless enterprise integration into commercial contact centers, audio narration engines, and digital companion products.

Benchmarks and Measured Results

According to independent evaluation by Artificial Analysis across 31 frontier audio models (Figure 1, Speech Leaderboard):

• Pronunciation Accuracy:
• Ming-3.0-tts achieved a chart-topping 93.2% score, ranking #1 globally among all 31 systems;
• Decisively surpasses commercial titan ElevenLabs v4 (91.7%), demonstrating superior accuracy across homographs, foreign loanwords, and technical phrasing;
• Controlled Voice Leaderboard:
• Scores 1,069 Elo points, clinching #1 among all open-weight models (#13 overall) ahead of Drama 3 Preview (1,059);
• Provider Voice Leaderboard:
• Reaches 1,188 Elo points, securing #3 among open-weights (#19 overall).

Getting Started for Developers

• Open Weights: Model checkpoints and inference harnesses are available on Hugging Face under Ming-3.0-tts;
• On-Device Deployment Recommendation: The 0.5B dense variant runs with under 2GB of VRAM, making it ideal for pairing with local LLMs (e.g., Qwen3.5-4B) to build fully offline, zero-latency conversational voice agents without external API dependencies.