Close Menu
  • AI
  • Content Creation
  • Tech
  • Robotics
AI-trends.todayAI-trends.today
  • AI
  • Content Creation
  • Tech
  • Robotics
Trending
  • Banks Have Had It As much as Right here With Your AI Agent
  • Greatest Private AI Brokers in 2026: OpenClaw, Hermes Agent, Gemini Spark and Extra, In contrast on Reminiscence, Safety and Price
  • CrowdStrike Reportedly in Talks to Purchase AI Safety Associate Zscaler – Unite.AI
  • Tesla Rolls Out FSD V15 with Actuality View and Imaginative and prescient Upgrades – Unite.AI
  • Potential Cognition AI Spherical Would Elevate $2B-Plus at $75B Valuation – Unite.AI
  • OpenRouter Cuts Three Suppliers After AI Immediate Publicity Findings – Unite.AI
  • A Group of Researchers from Columbia and CertiK Introduce TensorLean for Formal Verification of GPU Kernels
  • Collectively AI Discloses Immediate Leak After Coding Agent Uncovered Credential – Unite.AI
AI-trends.todayAI-trends.today
Home»Tech»Gradium AI releases new default TTS model: 81.0% hard-case pass rate at 216ms time-to-first-audio

Gradium AI releases new default TTS model: 81.0% hard-case pass rate at 216ms time-to-first-audio

Tech By Gavin Wallace01/09/20265 Mins Read
Facebook Twitter LinkedIn Email
Step-by-Step Guide to Creating Synthetic Data Using the Synthetic Data
Step-by-Step Guide to Creating Synthetic Data Using the Synthetic Data
Share
Facebook Twitter LinkedIn Email

Voice agents often fail to deliver the information that is most critical: order numbers, callbacks, and email addresses. Gradium AI’s Studio and API now use a default text-to-speech engine. It reports an 81.0% pass rate human-rated on a set of 500 sentences in five different languages. This is ahead of Cartesia Sonic 3.6.1 at 75.1%, and ElevenLabs Conversational v3 at 65.4%. Coval has a time to first sound of 216 ms (P50) which is 170 ms more than the previous model.

Is it deployable?

Today, yes, without migration. Gradient switched the model on The API is set to default in all its applications. Studio The deadline for this is August 31st, 2026. Existing voices, including custom clones, keep working unchanged.

Accuracy number

Gradium created a set of 500 sentences for evaluation. open-sourced it on Hugging Face Under CC BY 2.0: 100 items in 10 categories across five languages (EN DE FR ES PT). Seven atomic criteria are based on spelling, acronyms and alphanumeric symbols, as well as dates, numbers (regular and large), floating numbers and emails. Three composite criterions (Orders Ticket IT, Claims), combine several criteria into one real agent turn.

The scoring is strict and human. The sentence is only passed if a native speaker rater independently hears all elements pronounced properly and fully; if a digit is dropped, the sentence fails. The raters were limited to 40 comparisons, with a forced break, after the audio had been loudness normalized and randomized.

All ten criteria are pooled together and the average is calculated over all five languages. The results were generated using default settings in August 2026.

Latency is measured by a number

You can find out more about this by clicking here. Coval’s TTS benchmarkGradium has reported a P50 time of 216 ms to the first audio. This is 170 ms more than its predecessor. A more meaningful figure is the range: 30 ms between the p75-p25 intervals across 480 runs. The tightest model of the five tested. Cartesia Sonic 3.6 has a 454 ms average with a 165 ms difference, or 36%, of the median. Its callers also experience tail-turns rather than medians.

The fastest model is Gradium, but it’s not even close. Gradium trails Inworld TTS 2, Fish Audio S2.1 pro (291 ms), and ElevenLabs Conversational (328 ms). This claim is made about the joint position. It claims to have the lowest failure rate for first audio at less than 250 ms, and with little variation.

Getting started

There is no need to update existing users. The software is installed by new teams. Python SDKUse existing voice IDs and point to the WebSocket TTS Endpoint. Gradium offers 1M credits to complete reports of hard-case failures on its Discord.

What you need to know

  • No migration required. The new Gradium TTS is default and live as of August 31st 2026.
  • Cartesia is the only software that has a pass rate of 81.0% on 500 sentences. ElevenLabs and Fish Audio are also in front.
  • Coval – 216ms to P50, first audio. 30ms spread between quarters.
  • It can read IBANs (International Bank Account Numbers), phone numbers and other codes without text normalization.
  • Hugging Face’s 500-sentence benchmark is available under CC BY 4.0.

Take a look at the release post You can also read about the dataset. Also, feel free to follow us on Twitter Don’t forget about our 150k+ML SubReddit Subscribe now our Newsletter. Wait! What? now you can join us on telegram as well.

You can partner with us to promote your GitHub Repository OR Hugging Page OR New Product Launch OR Webinar, etc.? Connect with us


Michal Sutter, a data scientist with a master’s degree in Data Science at the University of Padova. Michal Sutter excels in transforming large datasets to actionable insight. He has a strong foundation in statistics, machine learning and data engineering.

AI ar audio
Share. Facebook Twitter LinkedIn Email
Avatar
Gavin Wallace

Related Posts

Greatest Private AI Brokers in 2026: OpenClaw, Hermes Agent, Gemini Spark and Extra, In contrast on Reminiscence, Safety and Price

22/11/2026

A Group of Researchers from Columbia and CertiK Introduce TensorLean for Formal Verification of GPU Kernels

22/11/2026

OpenAI Ships GPT-Stay-1.1 for Voice Brokers, Cuts Idle Value to $0.02/Min

21/11/2026

Meet Pondering Machines Inkling 2 Mini: 26B Looped MoE Runs on a 32GB Mac

21/11/2026
Top News

Apple’s new Siri AI is ready to get personal

Now is the time of the all-access AI agent

OpenAI has just made a major math claim. Some Academics are Crying Foul

Mistral’s new ultra-fast translation model gives big AI labs a run for their money

AI Digital Twins Help People Manage Diabetes And Obesity

Load More
AI-Trends.Today

Your daily source of AI news and trends. Stay up to date with everything AI and automation!

X (Twitter) Instagram
Top Insights

The AI model can understand how the physical world works

07/12/2025

Stanford Researchers Launch TRACE: An Agentic Training System Targeted at Capabilities that Transforms Recurring Agent Failures to Synthetic RL Environment

13/07/2026
Latest News

Banks Have Had It As much as Right here With Your AI Agent

22/11/2026

Greatest Private AI Brokers in 2026: OpenClaw, Hermes Agent, Gemini Spark and Extra, In contrast on Reminiscence, Safety and Price

22/11/2026
X (Twitter) Instagram
  • Privacy Policy
  • Contact Us
  • Terms and Conditions
© 2026 AI-Trends.Today

Type above and press Enter to search. Press Esc to cancel.