Speech-to-Text (STT) Engineer – Fine-Tune ASR Model for Bahasa Indonesia

# Speech-to-Text (STT) Engineer – Fine-Tune ASR Model for Bahasa Indonesia ## About the Project We are building a production-grade AI voice platform and are looking for an experienced **Speech-to-Text (ASR/STT) Engineer** to improve transcription quality for **Bahasa Indonesia**. The objective is to fine-tune an existing open-source STT model using high-quality Indonesian speech datasets to achieve strong real-world performance across conversational speech, phone calls, and AI voice interactions. This is a contract/freelance role with the possibility of continued work for additional languages. ## Responsibilities * Fine-tune an existing open-source Speech-to-Text model (Whisper, Canary, Parakeet, NeMo, wav2vec2, etc.) * Identify, clean, and prepare high-quality Bahasa Indonesia datasets * Perform dataset normalization, text cleaning, and preprocessing * Improve recognition for: * Conversational Indonesian * Regional accents * Numbers * Dates & times * Currency * Email addresses * URLs * Proper nouns * Optimize inference accuracy while maintaining low latency * Evaluate model performance using WER/CER and qualitative testing * Document the training process so it can be reproduced ## Required Skills * Strong experience with Speech-to-Text / Automatic Speech Recognition (ASR) * Experience training or fine-tuning models such as: * Whisper * NVIDIA NeMo * wav2vec2 * MMS * Canary * Parakeet * Other modern ASR models * Experience working with Hugging Face * PyTorch * Audio preprocessing and dataset preparation * Knowledge of Indonesian language datasets * Experience with multilingual ASR is a plus ## Nice to Have * Experience with telephony audio (8kHz) * Experience building production AI voice systems * Knowledge of punctuation restoration * Speaker diarization * Confidence estimation * Forced alignment * Noise robustness ## Deliverables * Fine-tuned Bahasa Indonesia STT model * Training scripts * Dataset preparation scripts * Evaluation report (WER/CER) * Inference scripts * Documentation for deployment ## Please Answer These Questions 1. Which STT models have you fine-tuned before? 2. Have you trained models specifically for Bahasa Indonesia? 3. Which Indonesian datasets have you worked with? 4. What WER improvements have you achieved on previous projects? 5. Which GPU infrastructure do you normally use? 6. How would you approach improving recognition of conversational Indonesian? 7. Have you worked with telephony (8kHz) audio? 8. Please share GitHub repositories, research papers, demos, or previous work. ## Ideal Candidate We are looking for someone who has already trained speech models for low-resource or multilingual languages and understands the full ASR pipeline—from dataset preparation to model training and production deployment. Strong communication and the ability to independently recommend datasets, training strategies, and evaluation methods are highly valued. If you've successfully delivered high-quality ASR models in production, we'd love to hear from you.