Case Study: AssemblyAI achieves transparent, accelerated end-to-end PyTorch speech recognition development with Comet

Building an end-to-end Speech Recognition model in PyTorch with AssemblyAI

AssemblyAI built an end-to-end speech recognition model in PyTorch and faced the typical deep‑learning challenges of data preprocessing, model tuning, and experiment management while training a Deep Speech 2–inspired architecture. To log, visualize, and understand their model development pipeline, AssemblyAI integrated Comet (Comet.ml) into their training stack so they could centrally track experiments, code, and metrics during research and development.

Using Comet, AssemblyAI instrumented their training and evaluation scripts to log hyperparameters and metrics (loss, learning_rate, test_loss, CER and WER), model graphs, and comparisons across runs. Comet gave them a productive dashboard to compare experiments, monitor convergence and decoding performance, and iterate faster—improving team productivity and experiment reproducibility by providing per‑epoch, per‑run visibility into measurable metrics like CER and WER. Comet enabled AssemblyAI to more quickly optimize their model and workflows.

AssemblyAI

Michael Nguyen

Machine Learning Research