Audio AI Engineer

*]:pointer-events-auto R6Vx5W_threadScrollVars scroll-mb-[calc(var(--scroll-root-safe-area-inset-bottom,0px)+var(--thread-response-height))] scroll-mt-[calc(var(--header-height)+min(200px,max(70px,20svh)))]" data-turn-id="eb2fb562-0f4d-4169-91ef-d6e03992c560" data-turn-id-container="eb2fb562-0f4d-4169-91ef-d6e03992c560" data-testid="conversation-turn-8" data-turn="assistant"> Audio AI Engineer, #1085 Multilingual Speech-to-Text Engineer — On-Device Model Optimization, #1085 A Role with Purpose and Impact This role builds the speech recognition core of a mobile translation capability supporting a government agency's national security mission. The engineer will take large, high-quality speech-to-text models spanning many language families and adapt, compress, and optimize them so they run performantly on an iPhone — including handling the reality that speakers frequently mix in borrowed English terms mid-utterance, and the model needs to make a sound call on whether to transcribe those terms in English or in the source language's own transliteration. This is an applied ML role, not a research-only position. The strongest candidate can move fluidly from raw audio data, to model adaptation and compression experiments, to a rigorous evaluation framework — and can clearly explain what they're building, why it's better than the status quo, and how they'll know it worked. What This Role Is (and Isn't) This position owns the speech-to-text model — its data, its training/adaptation, its size and latency on-device, and its accuracy across languages. It does not own iOS application development, translation (source-language-to-target-language), or the Swift/AVFoundation integration layer; those are handled by a separate mobile engineering function this role will collaborate closely with. Key Responsibilities Data pipelines: Ingest, clean, segment, label, and version multilingual audio and transcript data, with attention to code-switching and borrowed-word phenomena across the target language set. Model adaptation: Fine-tune and compress large ASR models (using LoRA/QLoRA, quantization, distillation, or other parameter-efficient and size-reduction techniques as appropriate) to fit iPhone-class memory, latency, and battery constraints, while preserving transcription quality. Dynamic, per-language deployment: Design model packaging so language-specific weights can be selected and downloaded on demand based on use-case context (e.g., an operator interviewing a Chinese speaker pulls only the Chinese ASR weights). Loanword/transliteration handling: Build and evaluate model behavior for deciding when a borrowed English term should be transcribed as-is versus rendered in the source language's transliteration or native equivalent. Evaluation: Build reproducible evaluation pipelines (word/character error rate, latency, robustness to accent/noise/speaking rate/code-switching) and clearly articulate results against defined success criteria for each language and deployment target. Documentation & communication: Produce clear model cards, dataset documentation, and evaluation write-ups that let technical and non-technical stakeholders understand what the model does, how it compares to alternatives, and what its risks and limitations are. Required Qualifications Bachelor's degree in Computer Science, Data Science, Machine Learning, Computational Linguistics, or a closely related field. Strong data-engineering background building production pipelines for large, messy, or unstructured audio/text datasets. Hands-on experience fine-tuning or adapting speech/audio models using parameter-efficient methods (LoRA, QLoRA, adapters) and/or model compression techniques (quantization, distillation, pruning) for constrained hardware. Practical experience with ASR/speech-to-text model development and evaluation across multiple languages, including error analysis under real-world conditions (accents, noise, code-switching). Strong Python and SQL skills; experience wit…