SpaceXAI has introduced Grok Voice 3.0, the latest version of its speech recognition model, which ranks first in the Artificial Analysis Speech Model Index with a score of 86.9. This model is designed for use in businesses engaged in customer support and sales through voice agents. Easy to integrate, Grok Voice 3.0 operates solely via API, eliminating the possibility of open access to model weights and local versions.
Grok Voice 3.0 demonstrates significant performance improvements: it has a time to first sound of 0.81 seconds, which is 40% faster than GPT-Live-1 (Astra, medium), which has a result of 1.34 seconds. The cost of using the model is $0.08 per minute, which is also comparable to the previous version, Grok Voice 2.0, now considered obsolete. The model supports 28 built-in voices and automatic language detection for 20 different locales, making it adaptable to various users and situations.
In testing, Grok Voice 3.0 showed a high level of successful tasks—94.8%, and achieved a result of 98.5% in Big Bench Audio tests, placing it among the best offerings on the market. However, in criteria related to agent tasks, like τ-Voice, the model still lags behind GPT-Live-1 with a score of 71.6%. Nevertheless, Grok Voice 3.0 remains the highest in overall performance rating among available counterparts, thereby providing value for companies using voice agents to enhance their customer support. For production, developers are recommended to specify version grok-voice-3.0 instead of the alias grok-voice-latest.
This material is prepared solely for informational purposes and does not constitute financial advice or a recommendation.
Related articles
