AI Powered Voice Acting? - The FIRST LLM Designed for TTS

Hume AI’s Octave is being pitched as the first large language model built specifically for text-to-speech—one that doesn’t just read words, but follows acting-style instructions to deliver emotion, tone, and performance cues. In hands-on tests, the model can shift delivery mid-monologue (including sarcasm, frustration, and “voice trembles”) in a way that feels closer to voice acting than traditional TTS, where emotion is often limited to preset styles or post-processing.

The practical difference shows up in how Octave responds to script and direction. Instead of treating input as plain text to be converted into audio, Octave uses natural-language descriptions and “acting instructions” to steer performance. In demos, a beauty-blogger-style prompt produced a lively, expressive delivery that sounded like conversational narration, while a “monster in a cave” scenario only landed when the script and voice description matched closely—suggesting the model is sensitive to semantic alignment between what’s written and how it’s supposed to be performed.

Benchmarks cited by Hume place Octave ahead of 11 Labs in description-match and audio quality, with naturalness nearly tied. But the more revealing results came from live experimentation: Octave handled emotional transitions well, yet struggled with voice consistency. In a goblin monologue that moved from anger to exhaustion, the voice sometimes shifted pitch and character feel between paragraphs, making it sound like slightly different performers rather than one stable character. The tester tried tightening structure and refining acting prompts (e.g., “voice acting from an animated film”) to improve continuity, with partial improvement.

The workflow also matters. The “playground” encourages quick iteration, but the tester found that inserting stage directions in parentheses inside the script often led to them being treated as literal text. Moving those cues into a dedicated “acting instructions” field produced cleaner results. For longer or more controlled performances, the tester switched to “projects,” splitting the monologue into sections and assigning different acting directions per segment. That approach improved pacing and emotional nuance, though the underlying consistency issue still appeared.

A second character test—an old wizard ranting about modern magic—highlighted both strengths and limitations. Octave could sustain high-energy sarcasm and deliver the intended annoyance early on, but the energy sometimes sagged later in the playground version. Sectioning the dialogue into a project restored control and made the performance feel more coherent.

Overall, Octave is positioned as a strong fit for emotion-driven voice acting, character narration, and listen-first experiences where expressiveness matters more than perfectly stable timbre. For tasks like audiobook-style transcription or textbook reading—where consistent voice identity is critical—traditional TTS may still be preferable. Pricing is framed as competitive, with plans starting around $3 per month, and Octave offering a limited free credit allotment for early testing. Hume acknowledged the consistency problem and indicated it should improve in the coming weeks without requiring a new model version.