asr_eval：多参考与流式语音识别评估算法与工具

Speech recognition evaluation methodologies have undergone several significant enhancements. Initially, a novel string alignment algorithm is introduced, designed to facilitate multi-reference labeling, accommodate insertions of arbitrary lengths, and achieve superior word alignment accuracy. This innovation proves particularly beneficial for non-Latin languages, languages characterized by rich morphological structures, and for the precise annotation of noisy or long-form speech segments. Furthermore, a new benchmark dataset, DiverseSpeech-Ru, has been meticulously compiled. This dataset comprises extensive, in-the-wild Russian speech recordings, each meticulously annotated with multiple references. The objective of DiverseSpeech-Ru is to more accurately reflect the complexities and challenges inherent in real-world speech recognition scenarios, encompassing factors such as background noise, diverse accents, and non-standard pronunciations. Concurrently, novel evaluation metrics have been developed specifically for streaming speech recognition systems. These metrics offer a more precise assessment of system performance in real-time applications, considering critical aspects like latency, throughput, and the trade-off with recognition accuracy. These specialized metrics address unique challenges inherent in streaming systems, such as the timeliness of partial recognition results and the system's adaptability to continuously evolving input. Collectively, these advanced algorithms and tools furnish a more comprehensive and accurate framework for evaluating speech recognition systems, especially when dealing with intricate speech data and real-time applications. These innovations empower researchers and developers to more effectively identify and address bottlenecks in existing speech recognition systems, thereby fostering advancements within the field.