Ky-Ng/IPA-ASR

Data Constrained Phonetic-Based Automatic Speech Recognition (ASR) for Low Resource Languages # Low Resource Automatic Speech Recognition (ASR) ___ ## Introduction With the advances in Automatic Speech Recognition (ASR) for high-resource languages, this paper aims to open the door to understanding the magnitude of data needed for speech recognition tasks. In this paper, we compare **Wav2Vec2XLSR** (56K hours of data) against **Wav2Vec2IPA**, a fine-tuned Wav2Vec2-based architecture model trained on `evaluation/eval_model_xlsr.ipynb` Wav2Vec2IPA `evaluation/eval_model_w2v2ipa.ipynb` | ### Datasets - We expand on the existing HuggingFace `timit-asr/timit_asr` and contribute two HuggingFace datasets in this project: | Name | Description | | -------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------- | | `kylelovesllms/timit_asr` | Implementation of TIMIT dataset with Test and Train split | | `kylelovesllms/timit_asr_ipa` | Implementation of `kylelovesllms/timit_asr` with Validation Split and IPA Transcriptions | ### Models - Similar to `datasets`, the models used/trained in this project live in HuggingFace Hub | Models | Description | | ------------------------------------------------------------------------------------------------ | --------------------------------------------------------------- | | `facebook/XLSR-Wav2Vec2` | Baseline Evaluation Model | | `facebook/wav2vec2-base` | Pretrained Wav2Vec2 Model | | `Wav2Vec2IPA` and `Wav2Vec2IpaTokenizer` | Fine Tuned Model Wav2Vec2-Base model …