Zipf's Law Suggests a Three-Pronged Approach to Inclusive Speech Recognition

Explore Zipf's law and its implications for inclusive speech recognition in this 55-minute lecture by Mark Hasegawa-Johnson from the Center for Language & Speech Processing at JHU. Delve into the three types of words - frequent, infrequent, and out-of-vocabulary - and how speech recognition technology has evolved to address each category. Examine the power-law distribution in language demographics and its impact on speech recognition approaches. Learn about monolingual pre-training, multilingual knowledge transfer, and unsupervised ASR methods for languages with varying amounts of data. Discuss the challenges of speech recognition for individuals with disabilities and the importance of collaboration between researchers and affected communities. Gain insights from Hasegawa-Johnson's extensive research in speech production, perception, source separation, voice conversion, and low-resource automatic speech recognition.