Audio Algorithm Engineer

• Audio Algorithm Engineer (Master / PhD Holder)
• 5 Work Days Per Week
• Office Near Tai Seng MRT , Singapore
• Medical Benefits Provided
• Entitled to Yearly Bonus & Performance Bonus

Job Requirements

1. Master's Degree or PhD Holder in Computer Science, Electrical Engineering, Signal Processing, Acoustics, Artificial Intelligence or a related field, with a focus on speech or audio.
2. Hands-on experience training and evaluating deep learning models for audio, in signal-level tasks (such as speech enhancement or voice activity detection), understanding tasks (such as speech recognition, sound event detection or audio captioning), or both; able to pick up new audio tasks quickly.
3. Solid foundation in audio signal processing: STFT, filterbank and mel features, pitch and harmonics, room acoustics and reverberation, SNR and level.
4. Familiar with audio front-end processing (noise suppression, echo cancellation, dereverberation, gain control and voice activity detection), and with how each step affects listeners and downstream models.
5. Able to design and train small, streaming models under compute, memory and latency budgets; familiar with quantisation, pruning and knowledge distillation, and with measuring model size, MACs and real-time factor.
6. Familiar with audio understanding models: speech recognition (CTC, transducer, and attention or LLM-based decoders), sound event detection and audio classification, audio captioning, and pretrained audio encoders such as Whisper, BEATs or CLAP.
7. Familiar with large audio models: audio-language models, fine-tuning them (including parameter-efficient methods such as LoRA), and connecting audio models to LLMs for question answering, summarisation or dialogue.
8. Experience in speech is highly valued, especially multilingual speech recognition (including WER and CER scoring with text normalisation, for languages with and without spaces between words) and voice interaction (end-of-turn detection and turn-taking).
9. Experience with audio data and evaluation: augmentation and synthetic mixing (noise, room impulse responses, speed perturbation, SpecAugment), test sets split by speaker, device and room rather than only by clip, and the right metric for each task.
10. Strong programming skill in Python and PyTorch; familiar with Linux, multi-GPU training, experiment tracking and git.
11. Experience with far-field audio (for example the CHiME, DNS or AEC challenges) or on-device deployment (quantisation-aware training, ONNX / TFLite, sherpa-onnx or vendor NPU toolchains) is a plus.
12. Experience with voice agents (ASR, LLM and TTS cascades, or full-duplex speech models such as Moshi) or audio-visual models is a plus.

Job Responsibilities

• Speech Recognition: Develop streaming, multilingual speech recognition and keyword spotting for far-field audio from a single camera microphone. Adapt, fine-tune and compress open models (for example Whisper, Parakeet, Qwen3-ASR) and compare them on our audio.
• Voice Interaction: Develop voice activity detection, end-of-turn detection, turn-taking and barge-in (the user interrupts while the device speaks). Compare cascades (ASR, LLM, TTS) with full-duplex speech models.
• Speech Enhancement: Develop noise suppression, echo cancellation, dereverberation and target-speaker extraction for one microphone. Measure what each step does for the people who listen and for the models that come after it.
• Sound Events and Audio Understanding: Develop sound event detection, audio classification and audio captioning.
• Model Training and Evaluation: Run audio experiments end to end, from data preparation to training, fine-tuning and ablations. Build evaluation harnesses: WER and CER for each language, miss and false-alarm rates for detection, event-detection F1 or mAP, end-of-turn latency, PESQ, DNSMOS and echo-cancellation metrics.
• Slice the results by distance, level, SNR and room, and test on real camera recordings.
• Data Analysis and Preprocessing: Build audio data pipelines: speech and sound corpora (with commercial licences checked), synthetic mixing with room impulse responses and recorded camera noise, augmentation, and labels taken from the clean sources before mixing. Guide data labellers where needed.
• Collaboration and Deployment: Most of our models run on the camera itself. Design small, streaming, real-time models from the start, compress them (pruning, distillation, int8 quantisation), measure compute, memory and latency on camera chips, and score the quantised model, not only the float checkpoint. Work with firmware engineers to integrate the models into products.
• Research and Innovations: Stay up to date with speech and audio AI. Reproduce papers and baselines quickly, and test them on our data before we trust the published numbers.
• Documentation and Reporting: Write clear reports of experiments, with results that others can reproduce. Communicate findings to technical and non-technical stakeholders.