Wav2Vec2 Large XLSR Hindi

theainerd/Wav2Vec2-large-xlsr-hindi is an automatic speech recognition (ASR) model fine-tuned from facebook/wav2vec2-large-xlsr-53 for Hindi transcription. It is trained on the Multilingual and Code-switching ASR Challenges for low-resource Indian languages (MUCS) dataset and evaluated on the Hindi test split of Common Voice. The model requires input speech sampled at 16 kHz.

Key Strengths

• Built on the pretrained XLSR-53 architecture, which leverages cross-lingual representation learning from 53 languages.
• Fine-tuned for Hindi using the MUCS challenge corpus (95.05 hours of speech from stories domain).
• Directly usable without an external language model; outputs character-level transcriptions.

Benchmark Performance

On the Common Voice Hindi test set, the model achieves a Word Error Rate (WER) of 72.62%. For context, the MUCS 2021 challenge baseline on the MUCS Hindi test set was 37.2% WER, while the top entries reached 12–14% WER. Differences in evaluation datasets and preprocessing mean these figures are not directly comparable.

Licensing

The base model (facebook/wav2vec2-large-xlsr-53) is released under the Apache-2.0 license. The fine-tuned Hindi model follows the same license.

Training Details

The fine-tuning script is publicly available as a Colab notebook. No external language model is used during inference.