IndicConformer Bengali 120M
The Bengali release is a language-specific IndicConformer recognizer. Its published encoder has 120 million parameters and a hybrid CTC–RNNT decoder.
Restricted access
The checkpoint is published specifically for Bengali
The inference instructions use AI4Bharat's NeMo fork and demonstrate both CTC and RNNT decoding. Audio preprocessing converts the recording to mono at 16,000 Hz. The 120M figure describes the encoder, not a separately measured total system size.IndicConformer Bengali 120M official model card
Keep the language-specific checkpoint with its decoder
Practical guidance. Use this release for a focused Bengali comparison. If recordings contain more than one language, compare it with the multilingual model before choosing the smaller encoder.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Restricted access |
| Access | Downloadable |
| Model identifiers | indicconformer_stt_bn_hybrid_ctc_rnnt_large |
| Encoder | 17 Conformer blocks; model width 512.IndicConformer Bengali 120M official model card |
| Published language | Bengali; repository language code bn.IndicConformer Bengali 120M official model card |
| Access | Gated NeMo checkpoint; MIT license.IndicConformer Bengali 120M official model card |
Related models
- IndicConformer 600M Multilingual
AI4Bharat · IndicConformer
600-million-parameter Conformer speech recognizer covering India's 22 scheduled languages.
Launch date not confirmedRestricted accessDownloadable