• 6 min read
Gradium’s new TTS model is live by default, claiming an 81% hard-case pass rate and 216ms median time to first audio.
Image: MarkTechPost Gradium has made a new text-to-speech model the default for its API and Studio. It targets failure cases that can make a voice agent unusable: spelling names, reading account numbers, handling acronyms and pronouncing email addresses without dropping characters.
The model went live on August 31, 2026. Existing users don’t need to migrate, and existing voice IDs—including custom voice clones—continue to work unchanged. Gradium says the model delivers a 216 ms P50 time to first audio on Coval’s TTS benchmark, while passing 81.0% of a 500-sentence, human-rated hard-case evaluation across English, German, French, Spanish and Portuguese.
A TTS system can sound natural in a demo yet fail when it has to read a callback number, a reference code or a dotted email address. Gradium’s claim is not that it has the lowest median latency in every test, but that it combines sub-250 ms first audio with a high success rate and a comparatively narrow latency spread.
“A voice agent giving you a tracking number has to be fast and correct: that’s table stakes. Speed and accuracy usually pull against each other. Our TTS model gives you both: it reads your hard cases correctly, in real time.”
The benchmark tests structured data, not just prose #
The evaluation set is available on Hugging Face under a CC BY 4.0 license. Its schema contains 500 rows, with 100 sentences per language and 10 published criteria. Seven are atomic checks: spelling, acronyms, alphanumeric tokens, dates, regular numbers, large and floating-point numbers, and email addresses. Three composite categories—Orders, IT Ticket and Claims—combine several of those elements in realistic customer-support turns.
The source dataset includes examples such as surnames dictated letter by letter, initials separated by periods, medication names, street names and strings containing acronyms such as FBI, NASA, IRS, EIN, SQL and HTML. The English examples also include phone and service numbers, email addresses, dates, booking references and flight numbers. That makes the set closer to the data-handling burden of a support call than a conventional read-aloud test, although it does not establish how the model performs on every production workload.
A sentence is counted as a pass only when a native-speaker rater hears every required element pronounced correctly and completely. One missing digit fails the entire sentence. Gradium says the audio was loudness-normalized, comparison order was randomized, raters were limited to 40 comparisons and two consecutive sessions, and a five-minute break was required to limit fatigue.
| Model | Hard-case pass rate |
|---|---|
| Gradium TTS | 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 scores were generated in August 2026 using default settings. Gradium says the overall result was pooled across all 10 criteria and averaged across the five languages with equal weighting. The evaluation is still vendor-produced: the dataset and methodology are public, but the supplied material does not include an independent replication of the ratings.
The company also says that the production path does not first rewrite input text with a language model or apply hidden text normalization. That matters for the comparison because a studio demo can otherwise be cleaner than the API output a developer receives. Gradium says the model accepts the original text and chooses how to pronounce structured character sequences itself, including phone numbers, email addresses, IBANs and reference codes.
“No rewriting, no hidden normalization, same input path as production, because the point of testing in the studio is to hear exactly what your users will hear.”
Latency is about the tail, not just P50 #
Gradium’s reported 216 ms P50 time to first audio is 170 ms faster than the model it replaces. It also reports a 30 ms p75–p25 interquartile spread across 480 runs, which the company says was the tightest among the five models tested. A narrow interquartile range doesn’t describe every tail-latency event, but it does indicate that the middle half of the measured turns clustered relatively close to the median.
| Model | Median time to first audio | Reported spread or comparison |
|---|---|---|
| Inworld TTS 2 | 166 ms | Fastest median listed |
| Gradium TTS | 216 ms | 30 ms p75–p25 spread across 480 runs |
| Fish Audio S2.1 Pro | 291 ms | Median listed on Coval |
| ElevenLabs v3 Conversational | 329 ms | Median listed on Coval |
| Cartesia Sonic 3.6 | 454 ms | 165 ms spread; 36% of its median |
The latency comparison uses model names and figures supplied in the release coverage; the hard-case accuracy table uses Inworld TTS 1.5 Max, while the latency figures identify Inworld TTS 2. Those are not interchangeable results. Gradium’s own comparison is therefore best read as a joint-position claim—high structured-data accuracy with relatively fast and stable first audio—not proof that it is the fastest system overall.
Coval’s benchmark documentation is said to publish the measurement point, regions, run counts and methodology, but those details are not reproduced in the supplied material. That leaves an important deployment question open: the 216 ms result may be useful for comparing the tested configurations, but it isn’t enough on its own to predict latency from a US production deployment with a particular network path, concurrency level or audio configuration.
No migration, but the release is still a benchmark claim #
Gradium opened a public beta for these hard cases on July 30, 2026, asking users to submit failures. The company says the new evaluation set was built from those reports. It is now available through the API and Studio, with the Python SDK connecting to Gradium’s WebSocket TTS endpoint. Gradium also says Pipecat and LiveKit are supported through direct integrations.
For existing applications, the operational change is small: the default model changes, while voice IDs and custom voices remain in place. New teams need a Gradium API key, the Python SDK and a WebSocket connection. The release material does not state a new price, a credit rate or a separate charge for the default model, so the cost of adopting it is unresolved from the information provided. Gradium is offering 1 million credits for complete hard-case failure reports submitted through its Discord. Submissions must include the input text and language, and the company says those examples will feed a future evaluation run.
The dataset makes the accuracy claim easier to inspect than a closed demo, but it doesn’t eliminate the limitations of a vendor benchmark. The 500 sentences are focused on customer-support failure modes, the ratings were commissioned by the model maker, and no independent benchmark result is supplied here. For teams building live voice agents, the release promises a default model available on August 31 with 216 ms median first audio, 30 ms middle-range spread, and an 81.0% pass rate on the specific structured-data set Gradium published.
Frequently asked questions #
Is Gradium’s new TTS model available now?+ #
Yes. Gradium made it the default across its API and Studio on August 31, 2026.
Do existing Gradium voices need migration?+ #
No. Existing voices, including custom voice clones, continue to work unchanged.
What did Gradium’s TTS model score?+ #
It recorded an 81.0% human-rated pass rate on a 500-sentence hard-case set covering five languages and 10 criteria.
How much does Gradium’s new TTS model cost?+ #
The supplied release material does not announce a new price or credit rate.
Ava Chen AI Editor
Ava covers the rapidly evolving world of artificial intelligence, from foundational models and research labs to the real-world economics of intelligence. With a background in computational linguistics, she cuts through the hype to find out what actually works. She firmly believes that benchmarks are just marketing until reproduced in the wild.