Tech Lead — ASR / TTS / Speech LLM (IC + Mentor)

* Prepare and maintain synthetic and real training datasets. * STT/TTS/Speech LLM and LLM training: model selection → fine-tuning → evaluation → deployment. * Build evaluation for clinical applications (RPM, triage, inbound/outbound). * Build scripts for data selection, augmentation (noise, codec, jitter), and corpus curation. * Fine-tune speech models using CTC/RNN-T or adapter-based recipes on multi-GPU systems. * Fine-tune LLMs using PEFT (LoRA/QLoRA/adapters) and preference methods such as DPO/RLHF where needed. * Implement evaluation pipelines to measure WER/sWER, entity F1, safety/quality metrics, and latency; automate MLflow logging. * Experiment with bias-aware training and context list conditioning. * Collaborate with backend and DevOps teams to integrate trained models into inference stacks. * Support creation of context biasing APIs and LM rescoring paths. * Assist in maintaining benchmarks versus commercial baselines (Deepgram, Elevenlabs, Cartesia, Whisper, etc.). * Optimize inference latency/cost for speech and LLM serving (batching, kv-cache, quantization, caching, autoscaling).

* Strong programming in Python (PyTorch, Hugging Face, NeMo, ESPnet). * Practical experience in audio data processing, augmentation, and ASR fine-tuning. * Training: SpecAugment, speed perturb, noise/RIRs, codec+PLC+jitter sims for PSTN/WebRTC. * Streaming ASR: Transducer/zipformer with chunked attention, frame-sync beam search, endpointing (VAD-EOU) tuning. * Context biasing: WFST boosts + neural re-scoring; patient/name dictionaries; session-aware bias refresh. * Familiarity with LoRA/QLoRA/adapters, distributed training, mixed precision. * Experience with LLM alignment and evaluation (SFT, DPO/RLHF, tool calling reliability, hallucination/safety checks). * Proficiency with evaluation frameworks: WER/sWER, Entity-F1, DER/JER, MOSNet/BVCC (TTS), PESQ/STOI (telephony), RTF/latency at P95/P99, and MLflow logging. * Inference/serving familiarity: vLLM/Triton, quantization, kv-cache, batching, and performance tuning. * Frameworks: ESPnet, SpeechBrain, NeMo, Kaldi/K2, Livekit, Pipecat, Dify. * Understanding of telephony speech characteristics, accents, and distortions. * Collaborative mindset for cross-functional work with ML-ops and QA.