About the Role
We are looking for an AI Expert to join our team as the technical anchor for next-generation real-time voice bot systems. While our Software Engineers focus on integration and infrastructure, this role owns the intelligence layer — designing, training, evaluating, and optimizing the AI models that make voice interactions actually work: from acoustic understanding to natural speech synthesis, speaker identification, and language comprehension in noisy, real-world conditions.
The ideal candidate has built voice AI systems in production — not just integrated APIs, but understood what's happening inside them and made deliberate choices about model architecture, training data, evaluation methodology, and latency trade-offs. You will work directly with Software Engineers to move models from research into production and with product stakeholders to define what good means for a voice bot.
Key Responsibilities
Voice AI Research & Development
- Design and develop end-to-end voice pipelines: ASR → NLU → dialogue management → TTS, with full understanding of the trade-offs at each stage
- Research, evaluate, and fine-tune models for Vietnamese-language voice tasks: ASR accuracy on domain-specific vocabulary (vehicle names, locations, plate numbers), TTS naturalness, speaker diarization in multi-party calls
- Build and maintain evaluation frameworks for voice AI quality: Word Error Rate (WER), Real-Time Factor (RTF), MOS scoring, speaker turn accuracy, end-to-end latency benchmarks
Model Fine-tuning & Training
- Fine-tune ASR models (Whisper, Gipformer, wav2vec 2.0) on domain-specific and accent-specific Vietnamese data
- Fine-tune TTS models (F5-TTS, VoxCPM2, VITS) to produce natural, brand-consistent voice output for CSKH scenarios
- Build and curate training datasets: audio collection pipelines, annotation tooling, data quality validation
- Apply speaker diarization and voice activity detection (VAD) to segment and attribute multi-speaker call audio
LLM & NLU Integration
- Design prompt strategies and context management for LLM-driven dialogue in voice-specific constraints (no markdown, short sentences, natural prosody cues)
- Develop intent classification, entity extraction, and slot-filling components tuned for spoken Vietnamese, including noisy transcripts from imperfect ASR
- Evaluate and mitigate hallucination in voice LLM pipelines, including ASR-grounded prompting and transcript correction techniques
Production & Optimization
- Optimize inference latency for real-time constraints: quantization, batching, streaming inference, model distillation
- Design and run A/B experiments on model variants in production; define statistical evaluation criteria for voice quality metrics
- Collaborate with Software Engineers on model serving architecture (GPU inference servers, ONNX runtime, TorchServe, Triton)
- Build monitoring pipelines for model performance drift, ASR degradation, and TTS quality regression in production
Required Qualifications
Education
- Bachelor's or Master's degree in Computer Science, Electrical Engineering, Applied Mathematics, Linguistics, or a directly related field. PhD is a plus for research-heavy responsibilities.
Hands-on voice AI experience (required)
- 3–5+ years of hands-on experience building and deploying voice AI systems — not just integrating cloud APIs, but working with the models themselves
- Demonstrated experience with at least two of: ASR fine-tuning, TTS model training, speaker diarization, VAD, voice conversion, or acoustic modeling
- Prior production experience with real-time voice pipelines where latency, reliability, and audio quality were hard constraints
Deep learning & NLP foundations
- Solid understanding of sequence modeling (Transformer, CTC, RNN), attention mechanisms, and how they apply to speech and language tasks
- Familiarity with acoustic features (MFCCs, mel spectrograms, filterbanks) and audio signal processing fundamentals
- Experience with speech-specific deep learning frameworks: ESPnet, SpeechBrain, NeMo, Hugging Face transformers + datasets for audio
Engineering skills
- Strong Python; comfortable with PyTorch or TensorFlow for model development and fine-tuning
- Ability to write production-quality code, not just research scripts: modular, testable, observable
- Experience with ML experiment tracking (MLflow, W&B), model versioning, and reproducibility practices
- Familiarity with Docker and GPU infrastructure for model serving
Evaluation mindset
- Knows that a model is only as good as its evaluation — can design rigorous offline and online evaluation protocols for voice AI
- Experience with human evaluation studies (MOS, MUSHRA, preference tests) and automated proxies
Nice to Have
- Experience with multilingual or low-resource language ASR/TTS, especially Vietnamese
- Familiarity with speaker diarization systems (pyannote, NeMo MSDD) applied to call center audio
- Knowledge of telephony audio constraints: G.711 µ-law/a-law codecs, 8kHz narrowband limitations, packet loss concealment
- Experience with streaming ASR (chunk-based inference, endpointing) for real-time applications
- Contributions to open-source speech/NLP projects or published research in voice AI
What This Role Is Not
- This is not an API integration role. We need someone who understands what happens inside the model, not just how to call it.
- This is not a pure research role. Models need to ship into production voice systems with real latency and reliability requirements.
- This is not a data annotation role. You will design the data strategy and evaluation criteria — others will execute annotation at scale.
What We Offer
- Ownership of the AI layer in a production voice system serving real customers
- Direct collaboration with Software Engineers, DevOps, and product teams — short feedback loops between research and production
- Access to real call audio data for training and evaluation (with appropriate privacy controls)
- Competitive compensation with equity
