AI Research Engineer- Speech 1

Summary

Centific is a frontier AI data foundry that empowers clients with innovative AI solutions. They are seeking an AI Research Engineer specializing in Speech/Audio to develop next-generation audio AI technologies, focusing on Large Audio Language Models and Speech-to-Speech systems.

Responsibilities

• Design, develop, and deploy Large Audio Language Models (LALMs) capable of native audio understanding, reasoning, and generation
• Build Large Audio Reasoning Models that perform complex chain-of-thought reasoning over speech and audio inputs, including medical, technical, and conversational domains
• Contribute to Speech-to-Speech (S2S) system development, including speech understanding, dialogue management, and speech synthesis components
• Research and implement alignment mechanisms between speech encoders and LLM backbones using lightweight adapters, LoRA, and efficient fine-tuning strategies
• Design efficient speech tokenization and temporal compression techniques suitable for long-form audio reasoning and multi-turn spoken dialogue
• Build comprehensive evaluation frameworks for audio reasoning capabilities, including benchmarks for speech QA, audio understanding, and reasoning accuracy
• Optimize inference pipelines for low-latency, streaming applications in speech systems
• Collaborate with cross-functional teams to transfer research innovations into production systems and customer-facing applications
• Contribute to technical documentation, research write-ups, and publications at top-tier venues (NeurIPS, ICML, ACL, Interspeech)

Skills

• Master's degree (required) or Ph.D. (preferred) in Computer Science, Electrical Engineering, or a related field with a focus on speech, audio ML, or multimodal learning
• 2+ years of industry or applied research experience in speech/audio AI, Large Language Models, or multimodal systems
• Demonstrated applied research contributions through publications, patents, or shipped products in speech/audio AI or LLMs
• Strong proficiency in Python and PyTorch, with hands-on experience in GPU-accelerated training for large-scale models
• Solid understanding of speech and audio signal processing, acoustic modeling, and audio representations
• Working knowledge of modern LLM architectures (Transformers, SSMs) and training paradigms including instruction tuning and alignment methods
• Familiarity with modality alignment techniques: adapter-based integration, cross-modal attention, or audio-text fusion methods
• Strong experimentation habits: clean code, systematic ablations, reproducibility, and clear technical communication
• Publication record at top-tier venues (NeurIPS, ICML, ICLR, ACL, Interspeech, ICASSP) in audio language models, speech reasoning, or multimodal learning
• Hands-on experience building or fine-tuning Large Audio Language Models (e.g., Qwen-Audio, SALMONN, LTU, Gemini Audio)
• Experience with speech representation pretraining (HuBERT, Wav2Vec 2.0, Whisper, WavLM) and discrete speech tokenization
• Familiarity with Speech-to-Speech components: neural audio codecs (EnCodec, SoundStream), vocoders, or speech synthesis systems
• Experience with audio reasoning benchmarks (AIR-Bench, MMAU, AudioBench) or building evaluation harnesses for audio QA
• Hands-on experience with distributed training (FSDP, DeepSpeed) and inference optimization (ONNX, TensorRT, quantization)
• Familiarity with speech frameworks such as ESPnet, SpeechBrain, NVIDIA NeMo, or Fairseq
• Experience with multilingual speech systems, code-switching, or domain adaptation for specialized applications (medical, legal, technical)
• Background in evaluating safety, bias, hallucination, or adversarial robustness in audio language models

Qualifications

Must Haves

• Master's degree (required) or Ph.D. (preferred) in Computer Science, Electrical Engineering, or a related field with a focus on speech, audio ML, or multimodal learning
• 2+ years of industry or applied research experience in speech/audio AI, Large Language Models, or multimodal systems
• Demonstrated applied research contributions through publications, patents, or shipped products in speech/audio AI or LLMs
• Strong proficiency in Python and PyTorch, with hands-on experience in GPU-accelerated training for large-scale models
• Solid understanding of speech and audio signal processing, acoustic modeling, and audio representations
• Working knowledge of modern LLM architectures (Transformers, SSMs) and training paradigms including instruction tuning and alignment methods
• Familiarity with modality alignment techniques: adapter-based integration, cross-modal attention, or audio-text fusion methods
• Strong experimentation habits: clean code, systematic ablations, reproducibility, and clear technical communication

Nice to Haves

• Publication record at top-tier venues (NeurIPS, ICML, ICLR, ACL, Interspeech, ICASSP) in audio language models, speech reasoning, or multimodal learning
• Hands-on experience building or fine-tuning Large Audio Language Models (e.g., Qwen-Audio, SALMONN, LTU, Gemini Audio)
• Experience with speech representation pretraining (HuBERT, Wav2Vec 2.0, Whisper, WavLM) and discrete speech tokenization
• Familiarity with Speech-to-Speech components: neural audio codecs (EnCodec, SoundStream), vocoders, or speech synthesis systems
• Experience with audio reasoning benchmarks (AIR-Bench, MMAU, AudioBench) or building evaluation harnesses for audio QA
• Hands-on experience with distributed training (FSDP, DeepSpeed) and inference optimization (ONNX, TensorRT, quantization)
• Familiarity with speech frameworks such as ESPnet, SpeechBrain, NVIDIA NeMo, or Fairseq
• Experience with multilingual speech systems, code-switching, or domain adaptation for specialized applications (medical, legal, technical)
• Background in evaluating safety, bias, hallucination, or adversarial robustness in audio language models

Benefits

• Competitive compensation package with comprehensive benefits
• Opportunity to work on cutting-edge Large Audio Language Models and audio reasoning research with real-world impact
• Collaboration with experienced applied scientists and engineers in speech and multimodal AI
• Support for publications at top-tier conferences and professional development
• Access to state-of-the-art GPU infrastructure for training large-scale audio models
• Flexible work arrangements with hybrid/remote options