Hume AI drops Octave 2 with sub-200ms voice generation

Most voice models are still struggling to sound like real humans without stumbling over complex words. Hume AI wants to go several steps further by giving users the ability to edit pronunciation at the phonetic level.

The emotion-focused AI company just introduced Octave 2, the second generation of its text-to-speech engine. The model is already rolling out in preview on Hume's platform and through its API.

Sub-200ms latency and 15-second clones

Octave 2 expands support to 11 languages, covering Arabic, English, French, German, Hindi, Italian, Japanese, Korean, Portuguese, Russian, and Spanish. Hume promised support for at least 20 languages over the coming months.

Cloning a voice now takes just a 15-second native speaker audio clip. When the cloned voice generates speech in another language, Octave 2 can automatically predict and apply the speaker's original accent.

Speed and pricing also got a serious upgrade. Octave 2 is 40% faster than its predecessor, generating responses in under 200 milliseconds. Hume teamed up with chipmaker SambaNova to build a custom inference stack for the architecture.

Pricing has been cut in half compared to Octave 1. For customers running dedicated deployments, Hume says the cost can drop to under one cent per minute of generated audio.

Editing phonemes like pixels

Hume is also introducing two upcoming features: realistic voice conversion and direct phoneme editing.

Voice conversion allows creators to swap one voice for another while locking in the exact timing, rhythm, and phonetic characteristics of the original performance. It is aimed directly at international dubbing and voice actor touch-ups.

Phoneme editing goes even deeper. Instead of regenerating entire paragraphs when a name is mispronounced, creators can adjust individual phonemes, tweak word emphasis, or even invent entirely new words from sound fragments.

Alongside the TTS launch, Hume released EVI 4 mini to bring these audio capabilities to its speech-to-speech API. It does not generate text natively yet, meaning developers still need to connect an external LLM.

Why it matters

For voice actors, game studios, and localization teams, text prompts have always been too blunt of a tool for professional audio work. Hume's approach of offering granular phoneme editing feels like the audio equivalent of introducing a brush tool to image generators, solving the constant headache of mispronounced brand names and weird inflections.

For the wider voice AI market, pushing latency below 200ms while driving dedicated costs below a cent per minute puts massive pressure on rivals like ElevenLabs and OpenAI. If interactive agents are going to handle customer support or gaming companion duties at scale, economics and speed will decide the winners.

That said, EVI 4 mini still requiring an external LLM shows that Hume has not quite cracked a fully unified end-to-end multimodal engine. Until phoneme editing officially ships and proves itself outside cherry-picked demos, creators will likely treat it with cautious optimism.