Indic-Transcribe Core
Core is the native-script member of the joint Indic-Transcribe family. It supports 25 languages, including Indian English, and can identify the spoken language.
Restricted access
Core keeps transcription in the native script
The supplied long-audio helper divides recordings into pieces of at most 30 seconds. The card warns that automatic language identification is uneven for Hindi-belt languages. It provides the same checkpoint in Hugging Face and NeMo formats; these are not separate models.Indic-Transcribe Core official model card
Pass a known language instead of asking the model to guess
Practical guidance. Core fits a workflow that stores transcripts in each language's own script. Check mixed-language names and spoken numbers before sending the transcript to a search or data-entry system.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Restricted access |
| Access | Downloadable |
| Model identifiers | indic-transcribe-core |
| Architecture | 1.2B parameters; FastConformer encoder and Transformer decoder derived from nvidia/canary-1b-v2.Indic-Transcribe Core official model card |
| Input | 16 kHz mono speech; designed for single-speaker audio.Indic-Transcribe Core official model card |
| Output | Native-script text. Spoken numbers remain words; English terms in Indic speech use the native script.Indic-Transcribe Core official model card |
| Access and terms | Gated weights under Indic Open Model License v1.0; upstream terms also apply.Indic Open Model License v1.0 |
Related models
- Indic-Transcribe Flex
Bodhan AI ยท Indic-Transcribe
1.2-billion-parameter speech-recognition model with native, mixed and Romanized transcript modes across 27 languages.
Launch date not confirmedRestricted accessDownloadable