ARTIFICIAL INTELLIGENCE (AI) RESEARCHER

Job Responsibilities

• Responsible for post-training work related to the speech understanding capabilities of Omni multimodal large models. Based on existing foundation models, design and conduct post-training experiments such as SFT, DPO, and RLHF to improve speech recognition, speech understanding, long-form speech reasoning, and speech instruction-following capabilities.

• Participate in the design, cleaning, construction, and quality evaluation of speech-text multimodal training data. Use Agents/tools where applicable to generate high-quality speech multimodal samples, identify bad cases, and drive dataset iteration and optimisation.

• Carry out algorithm iteration for key pain points in Omni model speech understanding, including noisy speech, accents, long-form speech, mixed speech, speech Q&A, and speech instruction understanding. Analyse model weaknesses, design solutions, and validate implementation results.

• Track frontier approaches in Omni models and speech large models, reproduce methods, conduct comparative experiments, develop practical algorithm solutions, prepare experiment reports, and align outcomes with business metrics.

• Work with inference, data, and engineering teams to complete model performance evaluation, build metric systems, and support online validation of optimised models.

Job Requirements

• Bachelor's degree or above in Computer Science, Speech Signal Processing, Artificial Intelligence, or a related field.

• 1-2 years of working experience in large models, speech algorithms, or related areas, with practical industry project implementation experience.

• Familiar with the full large model post-training workflow, including SFT, DPO, preference alignment, training data processing, prompt engineering, and model evaluation.

• Hands-on project experience in speech-side post-training for Omni or multimodal large models. Experience in speech instruction-following, long-form speech understanding, or speech dialogue alignment is preferred.

• Familiar with the basic principles of speech large models and Omni multimodal models, with knowledge of speech-text alignment and speech understanding technologies.

• Proficient in PyTorch and able to independently run post-training experiments.

• Strong bad case analysis skills, with the ability to identify issues from the perspectives of data, model, and training strategy, and drive model performance improvement.