Voice agents fail on precisely the parts of a call that matter most: the order number, the callback digits, the email address the caller has to write down. Gradium AI has released a new text-to-speech model and made it the default across its API and Studio, reporting an 81.0% human-rated pass rate on a 500-sentence hard-case set spanning five languages. That figure places it ahead of Cartesia Sonic 3.6 at 75.1% and ElevenLabs v3 Conversational at 65.4%, while the time to first audio measures 216 ms at P50 on the Coval benchmark, 170 ms faster than the model it replaces. The combination of aggressive latency and high accuracy on the edge cases that actually break customer interactions marks a substantive step forward for production-grade voice AI.
Why the hard-case pass rate matters more than average quality
Modern text-to-speech models have saturated performance on ordinary prose. Reading a news headline or a weather forecast is no longer a differentiator. The frontier has moved to the fragments of language that trip up statistical models: alphanumeric codes, spelled surnames, telephone numbers, email addresses, dates in ambiguous formats, and multi-digit reference strings. These are exactly the tokens that dominate support calls, banking verifications, travel bookings, and healthcare appointment confirmations. A TTS system that misreads “1-800-222-1222” as “one billion, eight hundred million” does not just sound unnatural — it destroys the utility of the call.
Gradium built a 500-sentence evaluation set and open-sourced it on Hugging Face under CC BY 4.0. The set contains 100 items across ten criteria in five languages (English, German, French, Spanish, Portuguese). Seven atomic criteria cover spelling, acronyms, alphanumeric tokens, dates, regular numbers, large and floating numbers, and email. Three composite criteria — orders, IT tickets, and claims — stack several of those into one realistic agent turn. Scoring is performed by independent native-speaker raters: audio loudness-normalized, order randomized, raters capped at 40 comparisons with an enforced break. A sentence passes only if every element is pronounced correctly and completely; one dropped digit fails the sentence. This strict protocol explains why the absolute numbers look lower than typical MOS scores — they measure correctness under real-world constraints, not subjective naturalness.
Pooled across the ten criteria and averaged over the five languages with equal weight, Gradium TTS achieved 81.0%. Cartesia Sonic 3.6 reached 75.1%, ElevenLabs v3 Conversational 65.4%, Fish Audio S2.1 Pro 49.5%, and Inworld TTS 1.5 Max 46.5%. All were generated in August 2026 with default settings. The gap between Gradium and the next-best competitor is six percentage points on a benchmark where a single mispronounced character counts as failure. For teams whose calls turn on a reference code — e-commerce support, telecom and IT helpdesks, claims intake, banking, travel, logistics — that difference can mean the difference between a resolved call and a repeat callback.
The latency picture: 216 ms median with unusually tight variance
On the Coval TTS benchmark, Gradium reports a P50 time to first audio of 216 ms across 480 runs. More telling than the median is the spread: the interquartile range (p75–p25) is only 30 ms, the tightest among the five models tested. Cartesia Sonic 3.6 sits at 454 ms median with a 165 ms spread — 36% of its own median. Callers experience tail turns rather than medians, so a wide variance means some users hear a response in 372 ms and others wait over 537 ms, creating an inconsistent user experience. Gradium’s narrow distribution implies that the model’s inference pipeline delivers predictable latency, which is critical for real-time conversational systems where jitter translates directly into awkward pauses or overlapping speech.
Gradium is not the absolute fastest model on that chart. Inworld TTS 2 posts a 166 ms median, and Fish Audio S2.1 Pro (291 ms) and ElevenLabs v3 Conversational (329 ms) trail Gradium. The claim being made is about joint position: the lowest hard-case failure rate at sub-250 ms first audio, with very little variance. For a voice agent handling phone calls, that combination means the agent can respond quickly and correctly even when the caller throws in an order number or an IBAN.
Deployability: live today with no migration
Gradium switched the new model on as the default across its API and Studio on August 31, 2026. Existing voices, including custom clones, keep working unchanged. New teams install the Python SDK, point at the WebSocket TTS endpoint, and reuse existing voice IDs. The model is hosted only — no open weights for this version, though on-device TTS remains a separate product line. Pipecat and LiveKit are supported directly, and the company provides a WebSocket TTS endpoint for custom integrations. The free tier offers 45,000 credits (roughly one hour of TTS at one credit per character), which allows teams to evaluate the model without upfront cost.
Gradium is also offering 1 million credits for complete hard-case failure reports submitted via its Discord. This bounty signals confidence in the model’s performance on the exact scenarios that matter most.
What the benchmark reveals about the state of TTS for voice agents
The Coval benchmark and the Gradium evaluation together paint a clear picture of where the industry stands in mid-2026. The traditional metric of mean opinion score (MOS) for naturalness is no longer the primary barrier. The remaining challenges are latency, consistency, and handling of structured data embedded in natural speech. The models that will win in production are not those that sound most human on a Wikipedia article, but those that can pronounce “AVN-1108-FT” or “[email protected]” without hesitation or error.
Gradium’s approach appears to treat the problem as a sequence of fine-grained tokenization and pronunciation decisions rather than a monolithic generation task. The 500-sentence evaluation set is designed to stress exactly those decisions. The fact that the model achieves 81% on a strict human-rated pass — while the next competitor is at 75% — suggests that the remaining errors are increasingly rare and perhaps concentrated in edge cases like overlapping digits in different languages or unusual email formats. The 19% failure rate still leaves room for improvement, but the gap to the field is substantial.
What is the “hard-case” set and why was it open-sourced?
The 500-sentence evaluation set was released under CC BY 4.0 to allow independent verification and comparative benchmarking. It is not a synthetic dataset generated by a language model. Each sentence was manually crafted to include at least one token that historically causes TTS errors: alphanumeric patterns like “ORD-4471-XB”, phone numbers, email addresses with dots, spelled-out names, and composite strings from order confirmations, ticket numbers, and insurance claim IDs. The ten criteria cover atomic types (spelling, acronyms, alphanumeric, dates, numbers, large/floating numbers, email) and composite types (orders, IT tickets, claims).
The decision to open-source the evaluation set is significant. It allows any team to run their own blind tests with raters of their choosing, reducing reliance on vendor-claimed numbers. Gradium published its own results alongside the dataset, but the community can now run head-to-head comparisons. This transparency builds trust and accelerates the entire field.
How the new model changes the cost-benefit equation for voice agent deployments
For engineering teams evaluating TTS for voice agents, the decision has typically involved a trade-off between latency and accuracy. Fast models tended to stumble on complex tokens, while accurate models added significant delay. Gradium’s result compresses that trade-off into a single region of the latency-accuracy scatter plot that was previously empty. The dot for Gradium sits at 216 ms and 19% failure rate, while the nearest competitor (Cartesia Sonic 3.6) is at 454 ms and 24.9% failure rate. The improvement is not incremental — it shifts the operational frontier.
This change has practical consequences for system architecture. Voice agent systems that previously used a fast TTS for general responses and a slower, more accurate model for specific token-heavy segments can now use a single model. Pipeline complexity drops, caching strategies simplify, and the user experience becomes more consistent. For contact-center operations that handle thousands of calls per day, reducing the average response time by 200 ms while simultaneously lowering error rates can improve both customer satisfaction and first-call resolution rates.
What the numbers mean for specific use cases
Consider a telecom support call where the agent needs to read back a customer’s phone number, account PIN, and a ticket reference. Under the Inworld TTS 1.5 Max model (46.5% pass rate), the chance that all three tokens are pronounced correctly is roughly 0.465^3 = 10% — in other words, 90% of such calls will contain at least one error. With Gradium TTS (81% pass rate), the compound probability rises to 0.81^3 = 53%. That is still far from perfect, but it represents a fivefold reduction in error probability per call. For an agent handling 200 such calls per day, the difference between 180 error-prone calls and 94 error-prone calls is substantial in terms of repeat callbacks and customer frustration.
The same arithmetic applies to banking: reading back an account number, a transaction amount, and a confirmation code. For travel: a booking reference, a flight number, a date. For logistics: a tracking ID, a delivery address, a customer name. Every domain where the call content is dominated by structured data benefits directly from hard-case accuracy.
Languages covered and what that means for global deployments
The published evaluation covers five languages: English, German, French, Spanish, and Portuguese. All languages were weighted equally in the pooled average, meaning the model’s performance is not tuned disproportionately to English. The 81.0% figure is a multi-language average. For teams running multilingual support centers, this eliminates the need to operate separate TTS engines per language. The same endpoint handles French alphanumeric tokens (like “AVN-1108-FT” pronounced in French) as well as Spanish or German ones. The dataset on Hugging Face includes language-specific sentences for each criteria, so teams can verify performance for their target languages independently.
Gradium’s announcement does not indicate immediate support for languages outside these five, but the architecture likely generalizes to other Latin-script languages. The company has not provided a roadmap for expansion, but the open evaluation set makes it straightforward for third parties to test model performance in additional languages by providing their own prompts.
Who should consider adopting Gradium TTS now
The primary buyers identified by the company are voice AI engineers, conversational AI product managers, contact-center operations leads, and machine learning engineers who want to re-run the open evaluation. The best-fit teams are those whose calls turn on a reference code — e-commerce support, telecom and IT helpdesks, claims intake, banking, travel, logistics. For teams that currently ship a TTS model that handles common sentences well but struggles on order numbers or email addresses, the upgrade path is immediate: no migration, no voice retraining, just a new default that scores 81% instead of 65% or lower.
For teams that have already invested heavily in custom voice clones or fine-tuning, the model preserves those assets. The same voice IDs continue to work. The only change is the underlying text-to-speech engine, which now handles the hard cases better without any additional configuration.
The 1-million-credit bounty for failure reports is an unusual but effective mechanism. It incentivizes developers to push the model to its limits and report back, generating a community-driven error corpus that Gradium can use for the next iteration. Teams that participate effectively gain free TTS credits and contribute to model improvement.
The broader market context: TTS as the bottleneck for conversational AI
Voice agents have seen rapid improvement in natural language understanding and dialogue management over the past three years. The weak link has increasingly become the voice output: if the agent understands perfectly but then mispronounces the critical piece of information, the call fails. Gradium’s release addresses that bottleneck directly. The 81% hard-case pass rate is not a final target — it is a benchmark that will likely rise in subsequent versions. But the fact that a model now exists at 216 ms with that accuracy changes the baseline for every competitor. Teams evaluating TTS for voice agents will now ask: “Can you read an IBAN correctly in 250 ms?” rather than “How natural does your model sound on a paragraph from Wikipedia?”
The open evaluation set creates a standard for answering that question. Any vendor can now run the same 500 sentences, with the same scoring rubric, and publish their numbers. The result will be a transparent leaderboard for exactly the use cases that matter. Gradium has effectively raised the bar for the entire industry by defining what “good” means in practical terms.
What to watch for next
Gradium’s next milestones will likely include language expansion, lower latency (sub-200 ms), and further reduction in failure rate toward 90% and beyond. The narrow latency variance suggests the model is already well-optimized at the inference level, so improvements may come from better tokenization of ambiguous strings, larger training datasets for rare patterns, or architectural changes that handle punctuation and spacing more robustly. The company’s investment in the Discord community and the open dataset indicates a commitment to iterative improvement driven by real-world usage rather than internal benchmarks alone.
For competing TTS vendors, the response will be telling. Cartesia, ElevenLabs, Fish Audio, and Inworld all have the evaluation set available now. If they can achieve comparable or better numbers within weeks, the market remains fluid. If Gradium holds a six-point lead after competitors train specifically on the hard-case set, that lead may indicate a genuine architectural advantage. Either way, the conversation about TTS quality has permanently shifted away from naturalness toward correctness under pressure.
That shift is long overdue. Voice agents that fail on the one piece of information the caller needs are not useful, no matter how natural they sound. Gradium has made correctness the headline metric, and the industry will be better for it.
