wav2vec2-large-xlsr-53-telugu

anuragshas/wav2vec2-large-xlsr-53-telugu is an automatic speech recognition (ASR) model fine-tuned for Telugu. It is based on Facebook’s XLSR-53 multilingual speech representation model and further trained on 70% of the OpenSLR SLR66 Telugu dataset. The model accepts 16 kHz mono audio input and produces transcribed text.

Key strengths

• Designed specifically for Telugu ASR, leveraging cross-lingual pretraining from 53 languages.
• Trained and evaluated on a standardized open dataset (OpenSLR SLR66), enabling reproducible comparison.
• Direct usage without a separate language model; output is generated via greedy decoding from the CTC head.

Best for

• Transcribing Telugu speech in applications such as media captioning, voice commands, and conversational analytics.
• Scenarios where a dedicated Telugu model is preferred over general multilingual alternatives.

Benchmark result

The model achieves a Word Error Rate (WER) of 44.98% on the test split of OpenSLR SLR66 Telugu.

Additional details

• Fine-tuned on Telugu only; not intended for other languages.
• Input audio must be resampled to 16 kHz for optimal performance.
• Text normalization (removal of punctuation, English characters, etc.) was applied during evaluation; similar preprocessing is recommended for production use.