New opportunity

Speech / Applied ML Engineer

VALSEA · Singapore

About this role

This fully remote, Singapore-based applied ML role focuses on improving speech performance for Southeast Asian languages, accents, code-switching and noisy audio. The engineer owns production-impacting model and pipeline improvements while balancing accuracy, robustness and latency against cost, memory and reliability constraints. The role calls for independent judgment with evolving evaluation criteria, imperfect multilingual data and incomplete labels, and an understanding of customer impact beyond BLEU and WER scores.

Responsibilities

Experiment with and tune speech and automatic speech recognition (ASR) models; develop approaches for multilingual and code-switching speech.
Design experiments under realistic production constraints, optimize inference and GPU utilization, and work with engineering to deploy model improvements.
Build evaluation suites and datasets, improve diagnostics and monitoring, investigate real-world failure modes, and refine data curation and annotation processes.
Document experiments, tradeoffs, notable failures and mitigation strategies. Own improvements for a speech use case or language and ship at least one measurable gain in accuracy, robustness or latency.

Required capabilities and working approach:

Solid Python and PyTorch fundamentals; understanding of speech and ASR; experience training, fine-tuning and evaluating models; and familiarity with GPU inference and optimization workflows.
Practical ML engineering judgment, initiative in proposing hypotheses or data strategies, and the ability to take experiments through to production improvements. Communicate uncertainties and risks clearly, learn from unsuccessful experiments, investigate root causes, and collaborate with researchers and engineers.

Preferred experience: Multilingual or low-resource speech, on-device or low-latency inference, and shipping ML models into production systems.

The posting describes direct collaboration with founders and senior engineers and opportunities to build a portfolio of production experiments and improvements. It does not specify a minimum number of years of experience or an education requirement.

For full details about this role, click “Apply on company site” to view the employer’s original posting.

Skills for this role

PythonPyTorchAutomatic speech recognitionSpeech model trainingModel fine-tuningModel evaluationGPU inferenceInference optimizationMultilingual speechCode-switchingExperiment designData curationData annotationProduction ML deploymentFailure analysis

Checking your profile…

Preparation suggestions, not the employer’s actual interview questions. Always check the original posting for current requirements.

Understand the role

What matters most

Ship a measurable speech improvement
The role owns model and pipeline changes for a speech use case or language, with success defined by a shipped gain in accuracy, robustness, or latency. Prepare to explain how an experiment could move from hypothesis to deployed, measured improvement.
Improve Southeast Asian speech performance
The posting emphasizes languages, accents, code-switching, and noisy audio. Prepare a concrete approach to tuning and evaluating ASR models across these conditions rather than relying on an aggregate score.
Design for production constraints
Experiments must account for latency, cost, memory, reliability, inference optimization, and GPU utilization. Prepare to explain how you would compare model gains with deployment costs and work with engineering on integration.
Build evidence and diagnose failures
The role includes evaluation suites and datasets, monitoring, root-cause investigation, and improvements to data curation and annotation. Prepare to show how you would document trade-offs, failed experiments, and mitigations while considering customer impact beyond WER or BLEU.

Likely challenges to think through

Measuring progress with imperfect data
Likely preparation challenge, not a confirmed employer problem: multilingual data and incomplete labels can make apparent gains unreliable or hide regressions in smaller speech slices. Prepare a plan for slice-level evaluation, label-quality checks, and stating uncertainty.
Balancing quality against serving costs
Likely preparation challenge, not a confirmed employer problem: a more accurate model may increase latency, GPU use, or memory. Prepare to describe how you would set realistic experiment constraints and assess whether a quality gain is worth its production cost.
Deciding when an experiment is ready to ship
Likely preparation challenge, not a confirmed employer problem: evolving criteria and unexpected production behavior can make a promising offline result difficult to deploy safely. Prepare to explain how you would communicate risks, investigate failures, and coordinate a measured change with engineers.

Practice your answers

Technical practice

How would you fine-tune and evaluate an ASR model for a Southeast Asian language or accent?

What to coverPreparation suggestion: Explain how you would establish a baseline, prepare speech and transcripts, run controlled training or fine-tuning in PyTorch, and compare results on held-out data. Address overfitting and whether the improvement survives production latency and memory constraints.

How would you improve recognition of multilingual and code-switching speech?

What to coverPreparation suggestion: Discuss how you would identify errors around language switches and compare model or data strategies against a baseline. Explain how you would check that gains in one language do not mask regressions in another.

How would you investigate poor ASR performance on accented or noisy audio?

What to coverPreparation suggestion: Describe how you would segment errors by accent and audio condition, inspect examples, and distinguish data problems from model limitations. Propose a targeted experiment and a robustness check rather than assuming a single fix.

How would you design a speech-model experiment under latency, cost and memory limits?

What to coverPreparation suggestion: State a hypothesis, baseline, comparison dataset and success criteria before running the experiment. Explain how you would record accuracy, robustness and resource trade-offs, including reasons to reject an offline gain.

How would you diagnose and improve GPU inference utilisation?

What to coverPreparation suggestion: Describe how you would measure latency, throughput, memory use and GPU utilisation before changing the inference workflow. Discuss possible bottlenecks and how you would test an optimisation for both performance and speech-quality regressions.

How would you curate speech data when multilingual labels are incomplete or inconsistent?

What to coverPreparation suggestion: Explain how you would inspect annotation quality, prioritise useful corrections and avoid treating uncertain labels as ground truth. Describe how you would keep dataset changes traceable so later model comparisons remain meaningful.

What would you include in an evaluation suite beyond aggregate WER or BLEU?

What to coverPreparation suggestion: Include slices for languages, accents, code-switching and noisy audio, alongside latency and robustness measures. Explain how you would relate errors to the speech use case and customer impact without claiming one aggregate score captures everything.

How would you investigate and monitor a real-world speech failure mode?

What to coverPreparation suggestion: Walk through reproducing the issue, checking audio and labels, comparing model versions, and testing a root-cause hypothesis. Explain what diagnostic slice or monitoring signal would help detect recurrence, plus how you would document the mitigation.

How would you move a promising ASR experiment into a production pipeline?

What to coverPreparation suggestion: Describe how you would work with engineers to verify pipeline compatibility, measure inference behaviour and check for regressions before deployment. Explain how you would document the change, its uncertainties and the measurable gain it is meant to deliver.

Behavioural practice

Tell me about a time you owned a model or pipeline improvement from idea to measurable outcome.

What to coverPreparation suggestion: Use situation, action and result to show how you chose a problem, tested an approach and followed through. Distinguish measured results from assumptions, including any production constraints.

Describe an experiment that failed and what you did next.

What to coverPreparation suggestion: Set out the hypothesis and result, then explain how you communicated the failure and investigated its cause. Close with what changed in your next experiment or decision.

Tell me about a time you incorporated conflicting feedback from researchers and engineers.

What to coverPreparation suggestion: Explain the situation, each perspective and the action you took to resolve a concrete trade-off. Describe the result and any uncertainty that remained.

Describe how you responded when a model behaved unexpectedly in production.

What to coverPreparation suggestion: Give the situation, your immediate actions to assess impact and risk, and how you coordinated the investigation. State the result and what you did to prevent or detect a recurrence.

Tell me about a decision you made with changing criteria or incomplete data.

What to coverPreparation suggestion: Explain what was unknown, how you chose a defensible next step and how you communicated the risk. Describe the result and how new evidence affected your decision.

Sharpen your knowledge

Topics to prepare

Speech and ASR fundamentals
The posting expects an understanding of speech and ASR basics. Revision exercise: explain how audio becomes a transcript, then identify where accent variation or background noise could cause errors.
Multilingual and code-switching speech
The role targets Southeast Asian languages, accents and code-switching. Revision exercise: sketch how you would compare errors across language and accent groups, including utterances that switch languages, without treating an aggregate score as sufficient.
Evaluation under imperfect data
The posting calls for evaluation suites and decisions with incomplete labels and evolving criteria. Revision exercise: outline a test set with slices for noisy audio and code-switching; note which conclusions would remain uncertain if labels were missing.
Production tradeoffs and impact
The role balances accuracy and robustness against latency, cost, memory and reliability, and asks for thinking beyond BLEU/WER. Revision exercise: compare two hypothetical model changes—one more accurate but slower, one faster but less robust—and explain what customer impact and production evidence would guide the choice.
Failure-mode reasoning
The posting emphasizes finding real-world edge cases and investigating root causes. Revision exercise: take a hypothetical cluster of transcription errors and list competing explanations involving audio quality, language coverage and labels, plus evidence that would distinguish them.

Skills to revise

Python and PyTorch model work
Solid Python and PyTorch fundamentals, including training and fine-tuning experience, are stated expectations. Practice writing a small, reproducible fine-tuning experiment for a speech model, recording the baseline, change and evaluation result; this is a preparation exercise, not a specified employer project.
ASR evaluation implementation
The role includes building evaluation suites and datasets. Practice writing Python code that compares reference and predicted transcripts, reports WER by relevant speech slice, and flags missing labels rather than silently counting them.
GPU inference measurement
Inference optimisation and GPU utilisation are explicit responsibilities. Practice measuring inference latency, memory use and GPU utilisation for a repeatable speech workload, then describe a possible optimisation and how you would check its effect on accuracy.
Data and error investigation
The posting calls for improved diagnostics, data curation and annotation processes. Practice reviewing a sample of hypothetical ASR errors, grouping them by likely cause, identifying questionable labels and proposing what should be re-annotated before another experiment.
Experiment-to-production communication
The role requires documenting tradeoffs and collaborating with engineering on production integration. Practice drafting a short experiment handoff with the measured gain, latency and cost implications, notable failures, remaining uncertainty and a monitoring plan. The handoff format is a preparation suggestion, not a stated company process.

Your preparation checklist

Prepare a speech-model walkthrough
The role calls for training, fine-tuning, and evaluating speech/ASR models. If you have a relevant project, prepare a brief walkthrough of the Python/PyTorch approach, baseline, experiments, results, and limitations; otherwise, label a proposed approach as hypothetical.
Sketch a multilingual ASR experiment
The posting focuses on Southeast Asian languages, accents, code-switching, and noisy audio. Prepare a hypothetical design with a testable hypothesis, data strategy, baseline, and criteria for deciding whether to continue.
Draft an evaluation and data plan
The role includes building evaluation suites and working with imperfect multilingual data and incomplete labels. Outline useful test slices, annotation or curation checks, diagnostics, and how you would interpret uncertain results beyond WER or BLEU.
Work through an inference tradeoff
The posting names GPU utilisation and latency, cost, memory, and reliability constraints. Prepare a short comparison of possible optimisation experiments, what you would measure, and when an accuracy gain would not justify deployment.
Rehearse a failure-to-deployment account
The role expects root-cause investigation, documented failures, and collaboration with engineering. Prepare one real example, if available, covering the failure, evidence, mitigation, deployment or monitoring decision, and what remained unknown; do not present a proposed case as shipped work.
Write questions about ownership and impact
The posting describes ownership of a speech use case or language and measurable production improvements. Ask which use cases need attention, how customer impact is assessed alongside speech metrics, what evaluation data is available, and how model changes reach production.

Availability last checked: 1 Oct 2026. Status can change after a check; confirm on the employer’s website. Report an issue · Our checking policy

KEEP EXPLORING

Similar AI jobs

Related skills and specializations in Singapore. Matched to this role, not your profile.

Explore more jobs

Finding similar opportunities…