A single multilingual model is superior to specialized single-language models for regional speech recognition