ML Researcher Developer for the Alignment team of the Speech Synthesis Service

ML Researcher Developer for the Alignment team of the Speech Synthesis Service

We are looking for a strong ML researcher developer for the Alignment team of the Speech Synthesis Service. Our goal is to bring synthesized speech to a level where it is absolutely indistinguishable from human speech in terms of intonation, emotionality, expressiveness, and naturalness.

Your developments will be the foundation that provides high-quality voice for Alice, the narration for Yandex Books, and the video translation technology in Yandex Browser. This is work with a huge impact on products used daily by millions.

Your key role is to design and implement advanced training methods to teach models to understand what it means to "speak well." We need an ML engineer-researcher ready to tackle complex scientific and technical challenges.

What tasks await you

Design of reward models You will create target metrics and reward models capable of accurately assessing the naturalness and expressiveness of synthesis, which will form the basis for fine-tuning.

Implementation of RL methods You will need to adapt and implement modern reinforcement learning algorithms (DPO, GRPO, and their modifications) to improve pronunciation accuracy, intonation variability, and speech naturalness.

Pipeline improvement Your task will be the continuous improvement of the core fine-tuning technology to increase the stability and efficiency of the process.

From idea to measurements You will be responsible for the full work cycle: from analyzing the latest scientific papers and conducting PoC to measuring the quality of the resulting model.

More about ML at Yandex — in the channel Yandex for ML

We expect you to

• Have a strong command of Python and possess expert experience with machine learning frameworks, particularly PyTorch
• Have practical experience with distributed training and large models
• Possess broad technical knowledge in NLP and are ready to implement new technologies from scratch
• Want to dive into the field of speech synthesis, are ready to understand both the theory and the engineering details of implementation
• Are capable of following developments in machine learning and turning research ideas into reliable, working code

It will be a plus if you

• Have worked with multimodal or generative models
• Have strong practical experience in reinforcement learning (RL)