New opportunity

Machine Learning Engineer (Speech/Audio) - Singapore

Plaud · Singapore

About this role

Plaud seeks an on-site Machine Learning Engineer (Speech/Audio) in Singapore to support speech recognition model development through large-scale data pipelines, model fine-tuning and evaluation. The role is part of its Global Product R&D Center.

Responsibilities

Own speech and audio data pipelines for collection, cleaning, filtering, labeling, augmentation and quality control at model-training scale. Work with senior speech engineers on mining terms and hotwords from ASR output.
Contribute to training, fine-tuning and evaluating speech/language models alongside domain specialists. Improve recognition of code-switching, names and product terms, and adapt models with scenario-specific data to recognize industry terminology across languages and business verticals.
Build keyword- and domain-lexicon-based test sets and evaluation frameworks; benchmark internal models against open-source and commercial alternatives.

Required qualifications

At least 1 year of hands-on experience in speech, machine learning or large-scale data engineering.
Experience in at least one of: ASR or SpeechLLM training, fine-tuning or evaluation; LLM or general machine learning model training or fine-tuning; or large-scale audio, video or text data pipelines, such as tens of thousands of hours of audio or TB-scale multimodal data.
Solid Python and PyTorch fundamentals, plus experience with distributed data processing, such as Spark or Ray.

Preferred qualifications

Familiarity with SpeechLLM or speech self-supervised learning, or exposure to StepAudio or Qwen3-Omni; experience improving ASR through hotwords, contextual biasing or code-switching; speech-related patents or publications at Interspeech, ICASSP or other leading AI venues; or ownership of a speech-data workstream spanning tens or hundreds of thousands of hours.

The full-time role offers market-competitive compensation, an employee stock ownership plan, AI productivity tools, equipment, company events, and medical insurance and WICA coverage. No salary amount is specified.

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

Skills for this role

PythonPyTorchSpeech recognitionASRSpeechLLMLarge language modelsMachine learningSpeech model fine-tuningModel evaluationAudio data pipelinesDistributed data processingSparkRayData labelingData augmentationData quality controlDomain adaptationCode-switchingHotword miningContextual biasingSpeech self-supervised learningStepAudioQwen3-Omni

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

Own training-scale speech data pipelines
Build and maintain collection, cleaning, filtering, labeling, augmentation and quality-control systems. Distributed processing is a stated requirement because the role owns pipelines at scale.
Improve recognition of difficult terms
Partner with senior speech engineers to mine terms and hotwords from ASR output, and contribute to model fine-tuning and evaluation focused on code-switching, names and product terms.
Adapt models to specific domains
Use scenario-specific data, alongside domain specialists, to improve recognition of industry terminology across key languages and business verticals.
Build evidence for model improvements
Create keyword- and domain-lexicon-based test sets and evaluation frameworks, then benchmark internal models against open-source and commercial alternatives.

Likely challenges to think through

Scaling data without diluting quality
Preparation hypothesis: More collected or augmented audio may increase training volume while introducing label errors or unsuitable examples. Prepare to explain how distributed processing, filtering and quality checks could work together at scale.
Targeted gains versus broader accuracy
Preparation hypothesis: Fine-tuning for names, hotwords or industry terms may help targeted cases without improving general recognition. Prepare to describe how you would compare targeted and broader test sets across languages and code-switching cases.
Making benchmarks comparable
Preparation hypothesis: Differences in model inputs, terminology support or test-set composition could obscure comparisons with external baselines. Prepare to explain a consistent evaluation setup and how you would report domain-specific results separately.

Practice your answers

Technical practice

How would you design a speech/audio data pipeline from collection through training-ready output?

What to coverSuggested answer focus: Explain collection, cleaning, filtering, labeling, augmentation and quality-control stages, and how data moves between them. Discuss traceability and checks that prevent unusable or duplicate examples from reaching training.

How would you detect and resolve labeling-quality problems in a large speech dataset?

What to coverSuggested answer focus: Describe checks for missing, inconsistent or poorly aligned transcripts, followed by targeted review of suspicious examples. Explain the trade-off between review effort and the quality of data retained.

When would you augment speech data, and how would you check whether it helped?

What to coverSuggested answer focus: Discuss matching augmentation to recognition errors rather than applying it indiscriminately. Compare performance on relevant held-out slices and check that augmentation has not harmed recognition elsewhere.

How would you use distributed processing for a speech-data pipeline at scale?

What to coverSuggested answer focus: Explain how you would partition work, handle failed tasks and monitor throughput using a distributed-processing approach such as Spark or Ray. Address skewed workloads and the trade-off between processing cost and repeatable outputs.

What Python and PyTorch fundamentals matter when fine-tuning and evaluating a speech or language model?

What to coverSuggested answer focus: Cover data loading, batching, gradient updates, training versus evaluation mode, and reproducible comparisons. Explain how you would diagnose a loss curve or evaluation result that looks inconsistent with the training setup.

How would you investigate poor recognition of code-switched speech?

What to coverSuggested answer focus: Break down errors by language combination and switching context before selecting data or model changes. Discuss how to test gains on code-switched examples without assuming improvements on other speech.

How would you mine candidate names or product terms from ASR output for hotword work?

What to coverSuggested answer focus: Describe how to find recurring term errors, validate candidate terms and prioritize them with senior speech engineers. Compare a hotword or contextual-biasing approach with adding training data, including the risk of false insertions.

How would you adapt a model to industry terminology across languages and scenarios?

What to coverSuggested answer focus: Explain how scenario-specific data could be selected and used for fine-tuning alongside domain specialists. Evaluate terminology gains by language and vertical while checking for regressions outside the target scenario.

How would you construct a keyword- and domain-lexicon-based ASR test set?

What to coverSuggested answer focus: Describe selecting representative terms, pronunciations or contexts and keeping evaluation examples separate from training data. Include both term-level results and broader recognition measures so a narrow improvement is not overstated.

How would you benchmark an internal recognition model against open-source and commercial alternatives?

What to coverSuggested answer focus: Use comparable inputs, transcription conventions and test slices for each system. Report differences for names, product terms, code-switching and domain terminology, noting limitations in the comparison rather than relying on one aggregate score.

Behavioural practice

Tell me about a time you took ownership of a data workstream with several processing stages.

What to coverSuggested answer focus: Use a real situation, your actions across the stages you owned, and the result. Clarify what you personally handled and how you verified data quality; do not imply a speech-data scale you have not worked at.

Describe a time you worked with a more experienced engineer to investigate model errors.

What to coverSuggested answer focus: Give the situation, how you shared evidence and incorporated feedback, and what changed as a result. This is suggested practice for the posting’s collaboration with senior speech engineers, not a claim about Plaud’s interview process.

Tell me about a time you translated a domain specialist’s input into a model or evaluation change.

What to coverSuggested answer focus: Explain the specialist’s concern, the action you took to make it testable, and the observed result. Distinguish your contribution from decisions made jointly with others.

Describe a time you found a data-quality issue that affected a machine-learning result.

What to coverSuggested answer focus: Set out how you noticed the issue, traced its cause, corrected or contained it, and checked the outcome. Choose a real example that shows how you handled uncertainty rather than claiming a particular speech project.

Tell me about a time you had to balance delivery speed against data or evaluation quality.

What to coverSuggested answer focus: Describe the constraint, the options you considered, the action you took, and the result. Explain what you measured or deferred so others could understand the trade-off.

Sharpen your knowledge

Topics to prepare

Speech-data quality across the pipeline
Posting: The role owns collection, cleaning, filtering, labeling, augmentation and quality control at model-training scale. Revision exercise (suggestion): For each stage, identify one likely failure and a check that would catch it.
Distributed processing at scale
Posting: Distributed data-processing experience is required; Spark and Ray are examples, not both required. Revision exercise (suggestion): Sketch how to process a large audio dataset in partitions and explain how you would detect duplicate work or failed partitions.
Model fine-tuning and domain adaptation
Posting: The role contributes to speech/language-model fine-tuning and uses scenario-specific data for industry terminology across languages. Revision exercise (suggestion): Explain how you would select adaptation data and compare a fine-tuned model with its starting model.
Recognition errors and hotwords
Posting: The focus includes code-switching, names and product terms, plus term and hotword mining from ASR output. Revision exercise (suggestion): Examine sample reference and ASR transcripts, group these errors, and propose which terms merit review as hotwords. Contextual biasing is preferred experience, not a required technique.
Evaluation and baseline comparison
Posting: The role builds keyword- and domain-lexicon-based test sets and benchmarks internal models against open-source and commercial alternatives. Revision exercise (suggestion): Design test subsets for names, product terms and code-switching, then specify how you would compare models on the same examples.

Skills to revise

Python data-pipeline implementation
Posting: Solid Python fundamentals and speech/audio pipeline ownership are expected. Revision exercise (suggestion): Write a small pipeline that validates records, filters unusable examples and reports counts at each stage.
Spark or Ray data processing
Posting: Distributed processing experience is required, with Spark and Ray given as examples. Revision exercise (suggestion): Using a distributed tool you know, partition a sample dataset, run a cleaning step and verify that output counts reconcile.
PyTorch model workflow
Posting: Solid PyTorch fundamentals are required, and the role supports model fine-tuning and evaluation. Revision exercise (suggestion): Walk through or implement a small training and evaluation loop, checking tensor shapes, loss calculation and the separation of training from evaluation data.
Keyword-focused test-set construction
Posting: The role builds keyword- and domain-lexicon-based evaluations. Revision exercise (suggestion): Create a small labeled test set with target terms and reference transcripts; compare two sets of model outputs and report both overall errors and missed target terms.

Your preparation checklist

Select evidence for the stated requirements
Prepare a brief evidence sheet covering Python, PyTorch, distributed processing such as Spark or Ray, and at least one experience path listed in the posting: speech-model work, general ML/LLM training, or large-scale data pipelines. Record your actual scope and results without implying experience you do not have.
Walk through a data-pipeline project
Prepare a diagram and five-minute account of a relevant pipeline, showing collection, cleaning, filtering, labeling, augmentation, quality checks, and any scaling decisions. Plaud lists ownership of these speech/audio pipeline stages as a responsibility; distinguish stages you owned from those you would propose.
Explain a model-training or fine-tuning example
Bring a concise account of a real PyTorch training, fine-tuning, or evaluation effort: data, objective, experiment choices, results, and limitations. Connect it to Plaud’s stated speech/language-model work where applicable, without presenting general ML experience as ASR experience.
Practice a hypothetical adaptation design
Sketch how you would use scenario-specific data and ASR outputs to improve recognition of code-switching, names, product terms, and industry terminology. Include a possible term-mining and hotword-review loop; label this as a proposed approach, not a Plaud process.
Draft an evaluation plan
Prepare a sample keyword- and domain-lexicon-based test-set design, including language or scenario slices, error analysis, and a fair comparison with open-source and commercial baselines. The posting calls for these evaluations; your plan is an interview exercise.
Prepare questions about the workstream
Write questions about current data volume and quality bottlenecks, priority languages and verticals, how domain specialists and senior speech engineers collaborate, and which evaluation results guide model changes. These are points to clarify with Plaud, not assumptions about its practices.

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…