Kyutai's Hibiki translates speech in real time with your own voice

Real-time voice translation usually strips away everything that makes you sound like you, replacing your personality with a robotic voiceover.

French AI lab Kyutai just introduced Hibiki, a simultaneous speech-to-speech translation model that currently handles French to English.

Translating in real-time without losing your voice

Hibiki generates both spoken and text translations on the fly as the speaker talks.

Instead of assigning a generic AI voice, the model preserves the speaker's original vocal identity throughout the translation.

It also adapts its speaking pace dynamically based on the semantic meaning of what is being said, aiming for a more natural cadence.

According to Kyutai, both objective benchmarks and human evaluations show that Hibiki outperforms previous systems in quality, naturalness, and voice similarity, bringing it closer to human interpreters.

Why it matters

Most speech translation systems feel clunky because chaining transcription, text translation, and text-to-speech creates latency and erases the speaker's tone.

Hibiki tackles speech-to-speech directly, proving that real-time translation does not have to cost you your own voice.

Right now, it is limited to a single direction between French and English, so it is more of an impressive technical milestone than an everyday tool for the masses.

Still, if this technology scales to more languages, it could fundamentally change international conferences and multilingual podcasts by making translated conversations feel genuinely human.