almugabo/kinyarwanda_asr

a repository with experimental scripts to finetune ASR models in Kinyarwanda # Kinyarwanda ASR a repository with experimental scripts to finetune ASR models in Kinyarwanda Automatic Speech Recognition (ASR) or Speech To Text (STT) is the task of transcribing spoken language automatically. ASR models for a given language are either (a) trained from scratch (b) created by finetuning existing ASR model in other languages Here an overview of different efforts to create Kinyarwanda ASR models ### (a) from scratch (1) NVIDIA ASR Kinyarwanda ASR using Mozilla Common Voice docs.nvidia.com (2) KINSPEAK Antoine Nzeyimana. 2023. Kinspeak: Improving speech recognition for kinyarwanda via semi-supervised learning methods. arXiv preprint (arxiv.org) (3) META - Omnilingual ASR a suite of models providing aASR capabilities for more than 1,600 languages ([ai.meta.com]) github.com paper: ai.meta.com ### (b) fine-tuned #### from the recent Digital Umuganda Leaderboard #### 1. w2v-bert-2.0-kinyarwanda-asr (first place Track A) Finetuned from: facebook/w2v-bert-2.0 badrex/w2v-bert-2.0-kinyarwanda-asr #### 2. Finetuned Whisper (first place Track B&C) used openai/whisper-large-v3 as a base model github.com !!! data correction : *we noticed that many of the examples seemed to have the wrong label, i.e. the text transcription seemed to be for a different audio file.* ##### others mbazaNLP/Whisper-Small-Kinyarwanda P.S: - NVIDIA has several tools to inspect the data (not clear if maintained) Speech data explorer docs.nvidia.com …