# A New Voice Synthesis Model Sets a New Standard for Accuracy in Support Calls
## When Every Digit Matters
Voice-powered agents have transformed customer service, but one persistent weakness keeps undermining their reliability: they consistently stumble on the exact information callers need most — order numbers, callback digits, email addresses, and reference codes. A synthetic voice that sounds natural but drops a single character from a tracking ID or flattens an email domain into a flat word renders the entire interaction useless.
A recently released text-to-speech model addresses this problem head-on. Developed by a AI company, the model prioritizes precision on structurally complex payloads — the kind of content that trips up conventional synthesis engines — and has made it the default option across its API and Studio platform. The company reports an 81.0% human-rated pass rate on a rigorous 500-sentence evaluation set, outperforming leading competitors including Cartesia Sonic 3.6 (75.1%) and ElevenLabs v3 Conversational (65.4%).
## How the Evaluation Was Conducted
The developers constructed a 500-sentence benchmark designed to stress-test text-to-speech systems on exactly the content that matters in real-world agent calls. The set is open-sourced and freely available for anyone to re-run.
The evaluation spans five languages — English, French, German, Spanish, and Portuguese — with 100 sentences each covering ten distinct criteria. Seven atomic criteria target specific linguistic challenges: spelling accuracy, acronym pronunciation, alphanumeric token handling, date formatting, regular numbers, large numbers, and floating-point numbers. Three composite criteria layer multiple atomic challenges into realistic agent interactions, including order processing, IT ticket handling, and claims intake scenarios.
The scoring methodology is deliberately strict and human-driven. Independent native-speaker raters evaluated audio outputs after loudness normalization and randomized playback order. Each rater was limited to 40 comparisons with enforced breaks to prevent fatigue. A sentence passes only if every single element — every digit, every letter, every punctuation mark — is pronounced correctly and completely. One missed character means the entire sentence fails.
## The Accuracy Results
Across all ten criteria and five languages, pooled with equal weighting, the results are as follows:
| Model | Human-Rated Pass Rate |
|—|—|
| New TTS Model | 81.0% |
| Cartesia Sonic 3.6 | 75.1% |
| ElevenLabs v3 Conversational | 65.4% |
| Fish Audio S2.1 Pro | 49.5% |
| Inworld TTS 1.5 Max | 46.5% |
The gap between the top performer and the rest is substantial. The new model maintains a roughly 6-percentage-point lead over the nearest competitor, and more than 34 percentage points ahead of the two lowest-ranked models. This difference is not marginal — in a support call where a single mispronounced digit can cause a failed transaction, it is the difference between a resolved call and a callback.
## Latency and Consistency
Accuracy means little if the voice arrives too slowly. On Coval’s TTS benchmark, the model delivers a median time to first audio of 216 milliseconds, which is 170 milliseconds faster than the model it replaced. The more revealing metric, however, is consistency. The interquartile range (p25 to p75) across 480 runs is just 30 milliseconds — the tightest spread among the five models tested.
For context, Cartesia Sonic 3.6 has a median of 454 milliseconds but a spread of 165 milliseconds — over a third of its own median. In a live conversation, callers experience the tail of the distribution, not the median. A model that is fast on average but unpredictable under load creates noticeable, frustrating delays. The new model’s narrow variance means callers experience a consistently responsive voice in nearly every interaction.
It is worth noting that while Inworld TTS 2 achieves a faster median of 166 milliseconds, it does so at a significant cost to accuracy, scoring well below the new model on the same hard-case evaluation set. The new model’s value proposition lies in its joint position: high accuracy combined with low latency and low variance.
## What the Model Handles Well
The model is specifically designed to handle the content types that break conventional voice agents:
– **Order references and tracking codes** — alphanumeric strings with dashes and mixed case
– **Email addresses** — preserving dots, @ symbols, and domain structure
– **Phone numbers** — reading digits individually, including emergency and service hotlines
– **Spelled names and acronyms** — pronouncing each letter separately when required
– **Hotel and booking references** — complex alphanumeric codes with letter-number combinations
– **IBANs and financial identifiers** — long strings of digits and letters with no normalization needed
No text preprocessing or special markup is required. The model ingests raw input and produces correct output directly.
## Getting Started
The model is available today as the default option across the company’s hosted API and Studio. No migration or configuration changes are required for existing users — all current voice profiles, including custom clones, continue to function without modification.
New teams can integrate quickly using the Python SDK, connect to the WebSocket TTS endpoint, and reuse existing voice identifiers. The platform supports direct integration with Pipecat and LiveKit frameworks. A free tier is available with 45,000 credits, which provides approximately one hour of text-to-speech generation at a rate of one credit per character.
The company is also offering one million credits to teams willing to submit comprehensive failure reports from the open evaluation dataset, incentivizing community-driven quality improvement.
## Who This Is For
The model is best suited for teams whose voice interactions depend on reference codes and precise information transfer. Primary use cases include e-commerce support, telecommunications and IT helpdesks, insurance claims processing, banking services, travel bookings, and logistics tracking. The ideal buyers are voice and AI engineers, conversational-AI product managers, contact center operations leads, and machine learning practitioners who wish to run their own evaluations on the open dataset.
## Frequently Asked Questions
**Q: Is the new TTS model available as open weights for local deployment?**
A: No. The model is currently available only through the hosted API and Studio platform. On-device text-to-speech remains a separate product line from the same company.
**Q: Do I need to change my existing voice configurations?**
A: No. The model was switched on as the default on August 31, 2026, with no changes required. All existing voices, including custom voice clones, continue to work as before.
**Q: What languages does the model support?**
A: The published evaluation covers English, French, German, Spanish, and Portuguese. The company has not yet published official support beyond these five languages, though the model may handle additional languages in practice.
**Q: Can I re-run the evaluation myself?**
A: Yes. The full 500-sentence evaluation set is open-sourced on Hugging Face under a CC BY 4.0 license. The company encourages independent re-evaluation and community benchmarking.
**Q: How does this compare to the fastest model on the latency chart?**
A: Inworld TTS 2 has a faster median time to first audio at 166 milliseconds. However, it scores significantly lower on the hard-case accuracy evaluation at 46.5%, and the new model offers much tighter latency consistency with only a 30-millisecond spread.
**Q: What integration frameworks are supported?**
A: The platform supports Pipecat and LiveKit directly, provides a Python SDK, and exposes a WebSocket TTS endpoint for custom integrations.
**Q: What is the cost for new users?**
A: A free tier is available with 45,000 credits, which equates to roughly one hour of text-to-speech generation at one credit per character. Additional pricing details are available through the company’s standard plans.
**Q: Why is the spread in latency more important than the median?**
A: In live conversations, callers perceive the worst-case experience, not the average. A model with a wide latency spread will occasionally produce noticeable delays that break conversational flow, even if its median response time looks fast. A narrow spread ensures consistent, predictable performance across every interaction.
## Conclusion
Voice agents have reached a point where naturalness alone is no longer sufficient. The content that matters most in customer interactions — reference codes, account numbers, email addresses, booking IDs — demands precision that generic synthesis models often fail to deliver. This new text-to-speech model addresses that gap directly, combining top-tier accuracy on structurally complex content with low latency and remarkable consistency.
The 81.0% pass rate on a strict, human-rated evaluation set across five languages sets a meaningful new benchmark. The open-source availability of the evaluation dataset invites the broader community to validate and push these results further. For teams building voice agents in support, finance, logistics, or any domain where a single misread digit can derail a conversation, this model represents a significant step forward.
The combination of zero migration overhead, broad framework support, and an accessible free tier makes it straightforward for teams to adopt and evaluate. As voice AI continues to mature, the focus is shifting from sounding human to being correct — and this release demonstrates that both goals can be achieved simultaneously.
Thank you for reading
