Hume AI provides speech and voice products built on its research into multimodal emotional expression. For builders the central product is Octave, a text-to-speech model that uses a language model to interpret the meaning of a passage and adjust pronunciation, pitch, tempo, and emphasis accordingly. Alongside it sit the Empathic Voice Interface (EVI) for speech-to-speech conversation, real-time expression measurement, and the Kairos platform for simulating and evaluating voice agents.
Octave voices are created in two ways: voice design from a natural-language description such as a role or personality, or voice cloning from around 15 seconds of audio. Requests carry up to 5,000 characters per utterance and return MP3, WAV, or PCM through streaming (HTTP or WebSocket) and non-streaming endpoints; an instant mode targets roughly 200 ms to first audio.
| Model | Languages | Model latency | Notes |
|---|---|---|---|
| Octave 1 | English, Spanish | ~200 ms | voice design is English-only |
| Octave 2 (preview) | 11 incl. Japanese, Korean, French, German, Russian, Hindi, Arabic | ~100 ms | word and phoneme timestamps, long-form context |
Features
- Voice design: generates a new voice from a text description of its character
- Voice cloning: builds a voice from a short audio sample
- Timestamps: Octave 2 returns word and phoneme timing for alignment and captions
- Streaming modes: HTTP and WebSocket streaming plus a non-streaming endpoint, with an instant low-latency mode
- Output formats: MP3, WAV, and PCM
- SDKs and tooling: TypeScript, Python, and .NET SDKs, a CLI, and an MCP server
- EVI: an empathic speech-to-speech interface, with EVI 4-mini as the current lightweight variant
- Evaluation: Kairos builds scenario suites, runs agent-to-agent and human-to-agent simulations, and tracks regressions; public voice AI leaderboards score models on human judgment