Speech-to-text (STT) performance is crucial for products relying on voice input, demanding rigorous benchmarking to ensure real-world accuracy and latency. Essential metrics include Word Error Rate (WER) and Word Accuracy Rate (WAR), though WER's effectiveness is limited as it treats all errors equally, overlooking the context in critical applications like healthcare and finance. Normalization discrepancies and biases in training data can distort results, while real-time transcription constraints often yield worse WER scores compared to asynchronous systems. Evaluating STT APIs requires testing with diverse, realistic audio samples, considering factors like background noise, accent diversity, and speaker variation to reflect production conditions accurately. Latency measures such as Time to First Byte (TTFB) and latency to final output are vital, with the latter being more indicative of real-world performance. Continuous monitoring and fine-tuning of STT systems are recommended to adapt to evolving user needs and maintain reliability, as exemplified by Gladia's Solaria model, which excels in challenging environments and offers broad language coverage with low latency.