Neil Zeghidour

@neilzegh

CEO @GradiumAI. Founder of @kyutai_labs. Invented neural codecs and audio LLMs. Prev. Google DeepMind/Brain, Meta, Toha Heavy Industries.

following
followers

Voice transfer across languages and accents is claimed to preserve distinctive timbre

A post says favorite voices can retain their unique sound when transferred across languages and accents.

2 Sources, 6d ago, first seen 6d ago

New voice catalog described as bigger, more diverse and more natural

A user quoted a Gradium AI post and called the new catalog more diverse and natural.

2 Sources, 13d ago, first seen 13d ago

Gradium's latest text-to-speech model reportedly leads Voice Arena's US English latency board

Gradium says the model's median time to first audio is under 100 milliseconds, down from 236 milliseconds for its previous model.

4 Sources, 14d ago, updated 14d ago

Goals, ideas and methods as starting points for research

A quote-post warns that “method-driven” research can become a trap: finding problems that fit a favored technique rather than focusing on the most important problems.

2 Sources, 30d ago, first seen 30d ago

A user shares a video as “GRPO rollouts on my poorly pretrained model”

The user describes the linked clip in terms of GRPO rollouts, calling the model “poorly pretrained.”

1 Source, 45d ago, first seen 45d ago

Matt Turck Retweets Neil Zegh on Models and Infrastructure

The retweet presents a statement on models and infrastructure as equal elements in value creation.

2 Sources, 47d ago, first seen 47d ago

Featured in

Voice transfer across languages and accents is claimed to preserve distinctive timbre

New voice catalog described as bigger, more diverse and more natural

Gradium's latest text-to-speech model reportedly leads Voice Arena's US English latency board

Goals, ideas and methods as starting points for research

A user shares a video as “GRPO rollouts on my poorly pretrained model”

Matt Turck Retweets Neil Zegh on Models and Infrastructure

Featured in

Voice transfer across languages and accents is claimed to preserve distinctive timbre

New voice catalog described as bigger, more diverse and more natural

Gradium's latest text-to-speech model reportedly leads Voice Arena's US English latency board

Goals, ideas and methods as starting points for research

A user shares a video as “GRPO rollouts on my poorly pretrained model”

Matt Turck Retweets Neil Zegh on Models and Infrastructure