AI News AI资讯 6h ago Updated 2h ago 更新于 2小时前 48

Gradium AI Releases New Default TTS Model: 81.0% Hard-Case Pass Rate at 216 ms Time-to-First-Audio Gradium AI发布新默认TTS模型:81.0%硬案例通过率,首音延迟仅216毫秒

Gradium AI released a new default TTS model (live as of August 31, 2026) achieving an 81.0% human-rated pass rate on a 500-sentence hard-case benchmark across five languages, outperforming Cartesia Sonic 3.6 (75.1%), ElevenLabs v3 Conversational (65.4%), Fish Audio S2.1 Pro (49.5%), and Inworld TTS 1.5 Max (46.5%) The model delivers a 216 ms P50 time to first audio on Coval's TTS benchmark with an exceptionally tight 30 ms interquartile range (p75-p25) across 480 runs, representing the lowest va Gradium AI发布新TTS模型,在500句困难案例测试中达到81.0%人工评分通过率,领先Cartesia Sonic 3.6(75.1%)和ElevenLabs v3 Conversational(65.4%) 首音频时间(TTFA)216ms(P50),比前代模型快170ms,且延迟方差极小(30ms IQR,480次运行) 测试集涵盖5种语言(EN/DE/FR/ES/PT)、10项标准(7项原子+3项复合),已在Hugging Face开源(CC BY 4.0) 支持零文本规范化直接读取电话号码、邮箱、IBAN、参考码等复杂token,无需额外预处理 2026年8月31日已作为默认

72
Hot 热度
68
Quality 质量
65
Impact 影响力

Analysis 深度分析

TL;DR

  • Gradium AI released a new default TTS model (live as of August 31, 2026) achieving an 81.0% human-rated pass rate on a 500-sentence hard-case benchmark across five languages, outperforming Cartesia Sonic 3.6 (75.1%), ElevenLabs v3 Conversational (65.4%), Fish Audio S2.1 Pro (49.5%), and Inworld TTS 1.5 Max (46.5%)
  • The model delivers a 216 ms P50 time to first audio on Coval's TTS benchmark with an exceptionally tight 30 ms interquartile range (p75-p25) across 480 runs, representing the lowest variance among tested models
  • The 500-sentence evaluation set is open-sourced on Hugging Face under CC BY 4.0, covering 10 criteria (7 atomic: spelling, acronyms, alphanumeric tokens, dates, regular numbers, large/floating numbers, emails; 3 composite: Orders, IT Ticket, Claims) across English, German, French, Spanish, and Portuguese
  • No migration is required for existing users; the model was switched on as default across Gradium's API and Studio, with existing voices and custom clones continuing to work unchanged
  • The model reads phone numbers, emails, IBANs, and reference codes with no text normalization required, directly addressing the failure points that matter most in voice agent calls

Why It Matters

Voice agents consistently fail on the most critical parts of a call—order numbers, callback digits, email addresses—making this benchmark directly relevant to anyone building production voice AI systems. The combination of high accuracy (81.0% pass rate) and low latency (216 ms P50) with minimal variance positions Gradium's model as a strong candidate for real-world deployment where both correctness and responsiveness are non-negotiable. The open-sourcing of the evaluation dataset also sets a potential new standard for transparent, human-rated TTS benchmarking in the industry.

Technical Details

  • Benchmark methodology: 500-sentence hard-case set with strict human rating—each sentence must pass all 10 criteria (7 atomic + 3 composite) to be counted as a pass; a single dropped digit fails the entire sentence. Audio was loudness-normalized, order randomized, and raters capped at 40 comparisons with enforced breaks to prevent fatigue.
  • Latency performance: 216 ms P50 time to first audio on Coval's benchmark, 170 ms faster than the previous default model. The interquartile range of 30 ms (p75-p25 across 480 runs) is the tightest among the five compared models, indicating highly consistent inference behavior.
  • Multilingual coverage: Evaluated across five languages (EN, DE, FR, ES, PT) with equal weighting in the pooled pass rate. The model handles alphanumeric tokens, IBANs, emails, and large/floating-point numbers without requiring text normalization preprocessing.
  • Deployment: Switched on as the default model across Gradium's API and Studio on August 31, 2026. Existing voice IDs and custom clones remain compatible with zero migration. New users can integrate via Python SDK and WebSocket TTS endpoint.
  • Evaluation transparency: The full 500-sentence dataset is open-sourced on Hugging Face under CC BY 4.0, and Gradium is offering 1M credits for complete hard-case failure reports submitted via Discord, encouraging community-driven validation.

Industry Insight

  • The emphasis on "hard-case" accuracy (numbers, emails, alphanumeric tokens) signals a shift in the voice AI market from naturalness-focused benchmarks to utility-focused ones—practitioners should prioritize models that handle structured data correctly over those that merely sound pleasant.
  • The tight latency variance (30 ms IQR) is as important as the median latency itself; callers experience tail turns, not medians, so models with low spread provide a more predictable and professional user experience in production voice agents.
  • Open-sourcing evaluation datasets (as Gradium did) may become a competitive differentiator, pushing the industry toward more transparent and comparable TTS benchmarks rather than proprietary, opaque scoring systems.

TL;DR

  • Gradium AI发布新TTS模型,在500句困难案例测试中达到81.0%人工评分通过率,领先Cartesia Sonic 3.6(75.1%)和ElevenLabs v3 Conversational(65.4%)
  • 首音频时间(TTFA)216ms(P50),比前代模型快170ms,且延迟方差极小(30ms IQR,480次运行)
  • 测试集涵盖5种语言(EN/DE/FR/ES/PT)、10项标准(7项原子+3项复合),已在Hugging Face开源(CC BY 4.0)
  • 支持零文本规范化直接读取电话号码、邮箱、IBAN、参考码等复杂token,无需额外预处理
  • 2026年8月31日已作为默认模型上线API和Studio,现有用户无需迁移,自定义声音克隆继续可用

为什么值得看

本文提供了当前语音代理场景下TTS模型准确率的横向对比数据,填补了"困难案例"(数字、邮箱、缩写等)评测的空白。对于构建客服、技术支持等语音Agent的团队,该模型在保持低延迟的同时显著降低关键信息误读风险,具有直接部署价值。

技术解析

  • 评测体系:500句困难案例集,100句×10项标准(拼写、缩写、字母数字token、日期、普通数字、大数字/浮点数、邮箱),3项复合标准(订单、IT工单、理赔)模拟真实Agent对话轮次。人工评分标准严格:一句中任一数字遗漏即判定失败。
  • 性能数据:Gradium TTS 81.0%通过率,对比Cartesia 75.1%、ElevenLabs 65.4%、Fish Audio 49.5%、Inworld 46.5%。TTFA 216ms(P50),IQR仅30ms,延迟稳定性显著优于竞品(Cartesia IQR 165ms,占中位数36%)。
  • 部署方式:2026年8月31日作为默认模型上线,现有声音ID和自定义克隆无需迁移。新用户通过Python SDK接入WebSocket TTS端点即可。
  • 开源与激励:评测集在Hugging Face开源(CC BY 4.0),Gradium提供100万积分激励社区提交完整失败案例报告。

行业启示

  • 语音Agent的"最后一公里"瓶颈正在被攻克:数字、邮箱、IBAN等关键信息的准确朗读是客服/技术支持场景的核心痛点,Gradium在保持低延迟的同时将困难案例通过率提升至81%,标志着TTS从"可用"向"可靠"演进。
  • 延迟稳定性比绝对速度更重要:30ms的IQR意味着用户几乎不会感知到"卡顿",而Cartesia等模型虽然中位数不低,但长尾延迟会导致用户体验骤降。未来评测应更关注P75-P25等分布指标而非仅看中位数。
  • 开源评测集将成为行业基准:Gradium开放500句困难案例集,有望推动TTS评测从"厂商自说自话"转向"可比对的标准测试",建议从业者关注并参与该评测体系的共建。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Speech 语音 Product Launch 产品发布 Evaluation 评测 Conversational AI 对话系统