AI4BharatIndicWav2Vec

IndicWav2Vec Language Releases

Nine original Fairseq recognition checkpoints remain accessible: Bengali, Gujarati, Hindi, Marathi, Nepali, Odia, Tamil, Telugu and Sinhala.

Acoustic models and language models are separate components

The project distributes dictionaries, lexicons and six-gram KenLM files separately.IndicWav2Vec original pretrained and recognition releases

Record which decoding resources are active

Practical guidance. The same acoustic checkpoint can produce different transcripts with an added language model. Compare complete decoding configurations rather than treating a format conversion as a new recognizer.

IndicWav2Vec Hindi Hugging Face

This published Hindi package ports an AI4Bharat Fairseq Wav2Vec2 recognizer to Hugging Face. It is a serving-format release, not evidence of a newly trained foundation model. The card says inference with an added language model is unsupported. It does not identify an exact original checkpoint checksum. The package declares Apache 2.0 and requires download approval.IndicWav2Vec Hindi Hugging Face official model card

IndicWav2Vec Odia Hugging Face

This published Odia package ports an AI4Bharat Fairseq Wav2Vec2 recognizer to Hugging Face. It is a serving-format release, not evidence of a newly trained foundation model. The Odia repository's text mistakenly repeats a Hindi heading. The repository identity is Odia, but the card does not identify an exact original checkpoint checksum. The package declares Apache 2.0 and requires download approval.IndicWav2Vec Odia Hugging Face official model card

Specifications & access

IndicWav2Vec Language Releases specifications
DetailValue and conditions
AccessDownloadable
Model identifiersindicwav2vec_v1_bengaliindicwav2vec_v1_gujaratiindicwav2vec_v1_hindiindicwav2vec_v1_marathiindicwav2vec_v1_nepaliindicwav2vec_v1_odiaindicwav2vec_v1_tamilindicwav2vec_v1_teluguindicwav2vec_v1_sinhalaindicwav2vec-hindiindicwav2vec-odia
TermsMIT.IndicWav2Vec original pretrained and recognition releases
Access caveatLater Kannada and Malayalam Kathbath checkpoint links returned access errors.IndicWav2Vec original pretrained and recognition releases

Related models

  • Original large-sized IndicWav2Vec speech-pretraining checkpoint, separate from language-specific recognizers.

    Launch date not confirmedDownloadable