Eleven v3 语音模型正式全量开放，符号与数字朗读错误率下降 68%

Eleven v3, our most advanced Text to Speech model, is now out of Alpha and generally available.

Since the Alpha release, we've continued refining the model. Two key improvements:

More stable. In testing, users preferred the new version 72% of the time over the previous Alpha release.

More accurate. We significantly improved how the model handles numbers, symbols, and specialized notation across languages.

Accuracy improvements

Text to Speech models need to interpret what you write and decide how to say it. The same symbols can mean different things in different contexts.

Consider a phone number: "+49 170 9876543"

In some cases, our models would read this as "plus forty-nine, one hundred seventy, nine million eight hundred seventy-six thousand five hundred forty-three" - interpreting the digits as large numbers rather than a digit sequence. The correct reading is "plus four nine, one seven zero, nine eight seven six five four three."

These kinds of errors showed up across categories - sports scores, chemical formulas, currencies, coordinates - anywhere the models had to interpret symbols and decide how to vocalize them.

We tested against an internal benchmark covering 27 categories across 8 languages.

Overall: 68% reduction in errors. Error rate dropped from 15.3% to 4.9%.

Error rate by category:

The improvements are most significant in categories where context determines interpretation -  where a colon might indicate a sports score, a time, or an aspect ratio depending on surrounding text.

Examples

Currencies — correct magnitude:

Input:  ¥250,000

Before: 25,000 yen

After:  250,000 yen

Chemical formulas — symbols preserved correctly:

Input:  SO₂

Before: "sulfur double" (garbled)

After:  "S O two"

Sports scores — context-aware interpretation:

Input:  Final score: 102-98

Before: "one hundred two minus ninety-eight"

After:  "one hundred two to ninety-eight"

Availability

Eleven v3 is now generally available across all platforms.