Tìm việc làm IT

1. Robot Speech Stack Architecture

• Define and own the end-to-end speech architecture for humanoid and service robots, covering audio front-end, wake word detection, VAD/endpointing, speaker verification/identification, and diarization.
• Drive edge–cloud deployment strategies (NVIDIA Jetson or equivalent) based on latency, power, memory, and privacy constraints.
• Evaluate, integrate, and optimize speech-to-speech architectures, including full-duplex/end-to-end speech LLMs (Moshi, GPT-4o Realtime, Qwen2.5-Omni/Qwen3-Omni, GLM-4-Voice, Mini-Omni, SpeechGPT, etc.) and cascaded streaming ASR–LLM–TTS pipelines.
• Make data-driven architectural decisions based on latency, barge-in performance, turn-taking, and Vietnamese speech robustness.

2. Wake Word, Voice Biometrics & Speaker Diarization

• Wake Word / Keyword Spotting (KWS): Develop low-power, on-device, bilingual Vietnamese–English wake word models with customizable wake words. Optimize false rejection rate (FRR) at fixed false alarms per hour (FA/hour) under challenging acoustic conditions, including public spaces, events, and mobile robot environments. Minimize false triggers caused by the robot's own speech through acoustic echo cancellation (AEC).
• Speaker Recognition & Voice Biometrics: Develop speaker verification and identification systems using ECAPA-TDNN, x-vectors, ResNet, and WavLM-based embeddings. Support short-duration enrollment, text-dependent and text-independent verification, and wake word–based speaker authentication. Design anti-spoofing and liveness detection mechanisms against replay attacks, synthetic speech, and voice-cloning deepfakes.
• Speaker Diarization: Build streaming and offline diarization systems using pyannote, NVIDIA NeMo/Sortformer, and end-to-end diarization approaches. Handle overlapping speech and enable real-time speaker attribution in multi-person robot interactions. Track performance using DER/JER under no-collar evaluation settings.

3. Data, Optimization & Deployment

• Design and build Vietnamese speech datasets covering regional accents, dialects, Vietnamese–English code-switching, children, elderly speakers, far-field speech, and real-world noise captured from deployment environments. Combine synthetic data generation (TTS and augmentation) with robust annotation and quality assurance pipelines.
• Establish internal benchmarks, evaluation protocols, and deployment go/no-go criteria using metrics such as FRR @ FA/hour, EER/minDCF, DER/JER, WER/CER, first-response latency, and barge-in success rate.
• Define privacy-preserving mechanisms for voice biometric data, including on-device processing where feasible, voiceprint encryption, user consent, data retention/deletion policies, and compliance with applicable Vietnamese personal data protection regulations.
• Provide technical leadership and mentorship to Speech Research Engineers and AI Engineers. Review system designs, research methodologies, and code quality.
• Collaborate closely with the Agentic AI team on agentic AI, on-device offline TTS, and conversational AI to ensure seamless integration between the speech stack and robot behavior.
• Drive research publications and patent applications where appropriate, translating research innovations into scalable, production-ready robotic systems.

• Master's or PhD degree in Computer Science, Signal Processing, Machine Learning, Electronics, or a related field.
• 8+ years of experience in speech/audio machine learning (or 5+ years with a PhD), including at least 2 years in a technical leadership role.
• Proven experience taking at least one speech AI system from development to production on real-world devices.
• Strong Python proficiency with hands-on PyTorch experience. Solid C++ skills for real-time audio processing and inference.
• Strong systems thinking and experimental design skills, including operating point selection, ROC/DET curve analysis, and error analysis across different operating conditions.
• Wake Word / KWS: Hands-on expertise in lightweight architectures such as CNN, CRNN, TC-ResNet, and streaming Transformers; negative mining strategies; FRR @ FA/hour evaluation; and always-on, low-power optimization.
• Speaker Recognition & Anti-Spoofing: Strong knowledge of ECAPA-TDNN, x-vectors, ResNet, and SSL-based embeddings (WavLM/HuBERT); AAM-Softmax, score normalization, calibration, ASVspoof benchmarks, and liveness detection.
• Speaker Diarization: Practical experience with pyannote, NVIDIA NeMo (MSDD/Sortformer), EEND/powerset, overlap-aware diarization, streaming diarization, and DER/JER evaluation.
• Speech-to-Speech & Real-Time Voice: Proven hands-on experience integrating at least one end-to-end/full-duplex speech LLM and one streaming cascaded ASR–LLM–TTS pipeline. Strong understanding of latency trade-offs, VAD/endpointing, barge-in, turn-taking, and acoustic echo.

• Research publications in Interspeech, ICASSP, ASRU, SLT, TASLP, or relevant speech/audio research at NeurIPS, ICLR, or ICML. Speech/voice-related patents are a strong plus.
• Practical experience with:
  • Microphone arrays, acoustic echo cancellation (AEC), beamforming, and on-device speech enhancement.
  • Speech toolkits and frameworks such as SpeechBrain, WeSpeaker, ESPnet, k2/icefall, NVIDIA NeMo, and Kaldi.
  • Vietnamese speech dataset development or other low-resource tonal languages.
• Knowledge of:
  • Privacy-preserving machine learning and regulatory requirements for biometric data processing.
  • Multimodal audio-visual speaker tracking and active speaker detection for robotics.

BENEFITS & COMPENSATIONS 

• Attractive income, competitive and commensurate with individual capabilities. 

• 13th-month salary, gifts on public holidays and special occasions, and performance-based bonuses. 

• Meal allowance, annual company trips, health insurance, and exclusive benefits within the Group’s ecosystem 

• Clear career development opportunities aligned with the company’s growth, with access to training programs based on capability and job role.