Case Study: Rime builds the world’s most realistic text-to-speech model with Weights & Biases

Rime powers 100 million phone calls with Weights & Biases

Rime, a company specializing in voice AI, faced the significant challenge of building a highly realistic and robust text-to-speech model for enterprise use. Their goal was to create a system that could handle complex, real-world requirements like domain-specific pronunciations and natural prosody, moving beyond the limited and often disliked voices common in customer service lines. To manage the intricate process of training such a model, they relied on Weights & Biases.

Using Weights & Biases Models, Rime meticulously logged all experiments, tracked loss curves, and compared audio samples to develop intuition and identify the best model checkpoints. Features like Sweeps for automated hyperparameter tuning and Custom Views for easy run comparisons were crucial for their small team's velocity. The solution resulted in Arcana, the world's most realistic spoken language model, which now powers over 100 million phone calls per month for major enterprise customers and has contributed to Rime recently doubling its customer base.

Rime