Octave 2 is a second-generation speech-language model (SLM) developed by Hume AI, designed for expressive and emotionally intelligent text-to-speech synthesis. Unlike traditional text-to-speech systems, Octave 2 is trained to understand how text informs the tune, rhythm, and timbre of acting, allowing it to generate speech that reflects nuanced emotional context—such as whispering or shouting—based on the underlying intent of the script.
Capabilities and Performance
The model provides multilingual support across 11 languages: Arabic, English, French, German, Hindi, Italian, Japanese, Korean, Portuguese, Russian, and Spanish. It features a 40% improvement in speed over its predecessor, achieving response latencies under 200 milliseconds. This efficiency was reached through a custom inference stack optimized for advanced large language model hardware, specifically designed for the model's SLM architecture.
Octave 2 introduces specialized features for speech manipulation, including voice conversion and direct phoneme editing. Voice conversion allows for swapping one voice for another while maintaining the phonetic timing of the original audio, while phoneme editing enables granular adjustments to pronunciation, emphasis, and pacing. The model also demonstrates increased reliability when pronouncing uncommon words, repeated terms, numbers, and symbols.