About Cantina:
Cantina Labs is a social AI company, developing a suite of advanced real-time models that push the boundaries of expression, personality, and realism. We bring characters to life, transforming how people tell stories, connect, and create. We build and power ecosystems. Cantina, our flagship social AI platform, is just the beginning.
If you're excited about the potential AI has to shape human creativity and social interactions, join us in building the future!
About the Role:
We’re looking for an ML Engineer focused on Data Quality to own the datasets that power our speech systems. You will be hands-on with audio and text data: auditing, denoising, filtering, labeling, and building the tooling and models that turn messy, large-scale data into reliable training corpora for TTS and adjacent tasks. You’ll develop data quality metrics and classifiers, run human-in-the-loop annotation programs, and integrate quality gates into our training and evaluation pipelines. Your work will directly improve model performance, robustness, and cost by driving the model ↔ data ↔ eval flywheel from the data side.
What You’ll Do:
Dataset ownership: define specs; audit and curate large-scale audio/text; close corpus gaps and fix sample-level issues.
Quality instrumentation: build automated gates/metrics (e.g., SNR, clipping, VAD, WER, SV/LID, safety) with dashboards; validate against listening tests.
Classifiers and filters: train lightweight models to tag, score, and filter data (VAD, ASR gating, LID, SV/diarization, noise/safety); calibrate to subjective outcomes.
Cleaning and integrity: apply denoise/dereverb/de-clip when beneficial; deduplicate and decontaminate; prevent leakage; maintain lineage and versioned releases.
Data selection: optimize mixtures via sampling, weighting, curriculum, and active learning; mine hard negatives and long-tail cases.
Tooling and pipelines: ship reproducible ETL and validation; integrate quality gates into training/eval; add monitoring and alerts.
Human-in-the-loop and compliance: run MTurk/vendor annotation with strong QC; ensure consent/licensing/policy compliance; collaborate across teams and document datasets.
What You’ll Bring:
Strong experience building ML-driven data quality systems for audio/speech, or equivalent data-centric ML experience with a track record of improving model outcomes via better data.
Proficient in Python and PyTorch; training/finetuning SSL-ASR (Whisper, Wav2Vec, BERT) models, CNN based classifiers and writing robust production code.
Audio/speech fundamentals: torchaudio/librosa/ffmpeg, spectrogram features (e.g., log-mel, MFCC), VAD/SAD, basic DSP, and audio QA.
Scalable data engineering skills: Spark/Beam or similar, SQL, Airflow or equivalent orchestration, and cloud storage/computing (AWS/GCP).
Familiarity with ASR/TTS metrics and tooling: WER, MOS/MOSNet, PESQ/STOI/ViSQOL, speaker verification (EER), diarization, language ID.
Experience with dataset validation, versioning, and experiment tracking; comfort debugging data issues from single samples to fleet-wide trends.
Ability to balance rigor with speed, and to translate ambiguous requirements into measurable data improvements.
Preferred Experience:
Shipped datasets and/or data quality tooling that moved the needle for TTS/ASR/VC in production.
Built and deployed classifiers for LID, SV/diarization, VAD, noise/glitch detection, or safety/content moderation for audio.
Ran crowdsourcing/vendor annotation at scale with strong quality control (honeypots, IAA, label aggregation).
Background in de-noising/enhancement and their effects on downstream TTS quality.
Contributions to open-source or publications in speech/audio/ML.
Experience with data governance, consent tracking, and policy enforcement.
