How does Hume Octave compare to other leading TTS models like Elevenlabs?

Our state-of-the-art speech-language model Octave is the first LLM explicitly designed for text-to-speech (TTS). Unlike traditional TTS systems, Octave understands the meaning and context of the content it generates, leading to more natural, fluid, and contextually appropriate speech.

But how does Octave stack up against other leading TTS models available today?

To answer this question, we conducted a study with 1,200 participants, comparing Octave against other prominent models using the same voices featured in Huggingface's popular TTS Arena. Each participant created their own unique prompts, ensuring diversity and realism reflective of actual user interactions.

Participants then listened to pairs of speech samples generated randomly by two competing models, rating them head-to-head in a blind evaluation. Every participant rated at least five pairs of speech samples.

In head-to-head matchups, Hume won 68% of the time. Elevenlabs came in second with a 60.9% win rate, and Papla in third with 54.9% win rate.

These results indicate that Octave produces higher-quality, more natural-sounding speech, even when handling short, isolated pieces of user-generated text. However, this study didn't even take advantage of Octave's greatest differentiator -- maintaining coherent and engaging speech across longer texts or narratives. For long-form content, such as audiobooks or interactive voice experiences with complex character dialogues and distinct emotional expressions, we anticipate Octave’s contextual intelligence to demonstrate an even greater advantage.

Hume remains committed to building models that can better anticipate users’ needs by deeply understanding and adapting to human expression. Explore Octave yourself at the TTS playground, and keep an eye out for future updates.