hhnigatu/ASR-via-Translitration

Code, data, and models for Zero-shot Amharic ASR with transliteration # Exploring Transliteration-Based Zero-Shot Transfer for Amharic ASR Hellina Hailu Nigatu and Hanan Aldarmaki. AfricaNLP 2025. paper ## Abstract The performance of Automatic Speech Recognition (ASR) depends on the availability of transcribed speech datasets—often scarce or non-existent for many of the world’s languages. This study investigates alternative strategies to bridge the data gap using zero-shot cross-lingual transfer, leveraging transliteration as a method to utilize data from other languages. We experiment with transliteration from various source languages and demonstrate ASR performance in a low-resourced language, Amharic. We find that source data that align with the character distribution of the test data achieve the best performance, regardless of language family. We also experiment with fine-tuning with minimal transcribed data in the target language. Our findings demonstrate that transliteration, particularly when combined with a strategic choice of source languages, is a viable approach for improving ASR in zero-shot and low-resourced settings. ## Models Links to models trained with transliterated data. ### Multilingual Models trained with data combined from two languages. Each language has 10k utterances, sampled to match the Arabic dataset. Spanish-Arabic * Spanish-Xhosa * French-Arabic * French-Xhosa * French-Spanish * Arabic-Xhosa ### Monolingual Models trained with data from a single transfer language, with transliterated transcripts. Arabic * Xhosa * French * Spanish ## Data The compiled dataset with transliterated transcripts can be found on Huggingface for Arabic, Xhosa, French, and Spanish. Below, we present the sources of our training and test data. For ALFFA dataset, we provide the manually reconstructed test transcripts in the Data folder of this repository. ## Training Data * **Arabic**: Kulkarni, A., Kulkarni, A., Shatnawi, S.A.M., Aldarmaki, H. (2023) ClArTTS: An Open-Source Classical Arabic Text-to-Speech Corpus. P …