CrisperWhisper – Speech to Text Model That Transcribes What You Say Verbatim (github.com)

CrisperWhisper 2.0 has launched as a groundbreaking speech-to-text model that enables precise, verbatim transcription of spoken language, distinguishing between what was said and what was meant. Unlike traditional systems, which often mix these two interpretations based on training data, CrisperWhisper offers explicit control over the transcription mode, allowing users to choose between a literal account of the speech, complete with disfluencies, and a clean, intended version. Additionally, it boasts impressive word-level timing accuracy, achieving a mean boundary error of just 30 ms for read speech, and superior performance across multiple languages. This model is significant for the AI/ML community as it addresses a critical need for accuracy in speech recognition applications, notably in fields like clinical speech analysis and content generation. With its capability to seamlessly handle long-form audio without introducing artifacts and the ability to upgrade existing transcripts with actual spoken disfluencies, CrisperWhisper serves as a catalyst for creating high-quality datasets. Furthermore, topping the Nyra Verbatim Speech Benchmark in terms of disfluency recognition highlights its potential to outshine existing systems, setting a new standard in multilingual speech-to-text technologies.