Gradium AI’s New TTS Model: 81% Hard-Case Pass Rate at 216 ms

On August 31, 2026, Gradium AI replaced its default text-to-speech model across its API and Studio with a new one that scores 81.0% on a human-rated 500-sentence hard-case benchmark covering five languages. The company says that puts it ahead of Cartesia Sonic 3.6 at 75.1%, ElevenLabs v3 Conversational at 65.4%, Fish Audio S2.1 Pro at 49.5%, and Inworld TTS 1.5 Max at 46.5%. Median time to first audio on Coval's benchmark is 216 ms, 170 ms faster than the model it replaces, with a 30 ms interquartile spread across 480 runs.

What happened

Data point	Value
Live date	August 31, 2026
Hard-case pass rate (Gradium)	81.0%
Benchmark size	500 sentences, 10 criteria, 5 languages
P50 time to first audio (Coval)	216 ms
Speed improvement vs. prior model	170 ms faster
Latency interquartile range (p75-p25)	30 ms over 480 runs
Evaluation dataset license	CC BY 4.0 on Hugging Face

Gradium AI swapped its default TTS model on August 31, 2026, with no action required from existing users. Custom voice clones carry over without any changes. New teams connect via the Python SDK and a WebSocket TTS endpoint, reusing existing voice IDs.

How the benchmark works

Gradium built a 500-sentence evaluation set and published it on Hugging Face under CC BY 4.0, making the test data publicly verifiable. The set covers 10 criteria across five languages: English, German, French, Spanish, and Portuguese.

Seven of those criteria are atomic: spelling, acronyms, alphanumeric tokens, dates, regular numbers, large and floating-point numbers, and email addresses. The remaining three are composite scenarios called Orders, IT Ticket, and Claims. Each composite stacks several atomic criteria into a single realistic agent turn, the kind a voice bot actually handles on a live call.

Scoring is strict and human-rated. A sentence passes only if an independent native-speaker rater hears every element correctly and completely. One dropped digit fails the whole sentence. Audio was loudness-normalized, order randomized, and raters were capped at 40 comparisons with an enforced break between sessions.

How do the five models compare?

Model	Hard-case pass rate	P50 time to first audio
Gradium TTS (new default)	81.0%	216 ms
Cartesia Sonic 3.6	75.1%	454 ms
ElevenLabs v3 Conversational	65.4%	329 ms
Fish Audio S2.1 Pro	49.5%	291 ms
Inworld TTS 1.5 Max	46.5%	166 ms

Inworld TTS 2 posts the fastest median at 166 ms, so Gradium is not the speed leader in isolation. What Gradium claims is a joint position: the lowest hard-case failure rate among models that also deliver sub-250 ms first audio. Cartesia Sonic 3.6, the closest accuracy competitor, sits at 454 ms median with a 165 ms latency spread, which works out to 36% of its own median. In practice, callers experience tail latency on many turns, not the median.

Why it matters

Voice agents break down most visibly on structured data: order numbers, callback phone numbers, IBANs, email addresses. A missed digit or mangled acronym forces a repeat, extends handle time, and erodes trust. Accuracy on these edge cases is where the real quality gap lives, not in reading plain prose.

A tight latency spread (30 ms interquartile range) matters separately from the median. Consistent response timing feels more natural in conversation; a model that is usually fast but occasionally stalls produces an uneven experience that is hard to diagnose and hard to fix at the application layer.

Teams building AI voice integrations now have a public benchmark dataset they can run themselves, which is a meaningful step up from trusting vendor-only numbers. The CC BY 4.0 license means anyone can reproduce or extend the test.

Our take

The benchmark is vendor-run, which is always worth noting. The saving grace here is that Gradium open-sourced the 500-sentence dataset, so the numbers are at least checkable. Independent replication on that Hugging Face dataset should happen before anyone makes a major vendor switch based purely on these figures.

The composite criteria (Orders, IT Ticket, Claims) are the most interesting part of the methodology. Testing realistic agent turns rather than isolated numbers or spellings is a better proxy for production failure modes. Most TTS benchmarks miss this entirely.

The 170 ms latency improvement over the previous Gradium model is the kind of gain that actually shows up in call quality scores, not just engineering dashboards. Sub-250 ms first audio keeps the conversation rhythm intact. For teams already on Gradium, this is a free upgrade with no migration cost. For teams evaluating voice stack options, the public dataset is worth running before committing. We cover the broader shift toward AI-led agent system design regularly on the Lumien news feed.

What to do about it

1. If you are an existing Gradium user, do nothing. The new model is already live as your default.
2. Download the 500-sentence evaluation set from Hugging Face and run it against your current TTS vendor to get a comparable baseline.
3. Test your specific hard cases: order numbers, IBANs, email addresses, and any alphanumeric reference codes your callers use.
4. Report complete hard-case failures to Gradium’s Discord channel to claim the 1M credits they are offering for verified failure submissions.
5. If you are evaluating voice stack options for a new build, contact us to discuss which TTS providers fit your latency and accuracy requirements before you commit.

Run the public benchmark on your own hard cases before switching vendors. The dataset exists now; use it.

Source: Marktechpost