AI4BharatMahaDhwani
MahaDhwani Pretrained Conformer
MahaDhwani Pretrained Conformer is a self-supervised speech encoder trained on data from 22 scheduled Indian languages. It returns acoustic embeddings rather than text.
The published model supplies an encoder
The release uses AI4Bharat's NeMo package. Its output is an encoder representation for a speech sample; a recognition or other task head is needed for a downstream system.MahaDhwani Pretrained Conformer official model card
Separate feature quality from transcript quality
Practical guidance. Evaluate whether its features help your downstream training task. The encoder alone cannot be judged by a transcript metric because it does not produce a transcript.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | MahaDhwani_pretrained_conformer |
| Architecture | 120M-parameter encoder; 17 Conformer blocks and model width 512.MahaDhwani Pretrained Conformer official model card |
| License | MIT.MahaDhwani Pretrained Conformer official model card |
| Audio | The example uses mono audio at 16,000 Hz.MahaDhwani Pretrained Conformer official model card |
Related models
- IndicConformer 600M Multilingual
AI4Bharat ยท IndicConformer
600-million-parameter Conformer speech recognizer covering India's 22 scheduled languages.
Launch date not confirmedRestricted accessDownloadable