Bhili ASR Nemotron 600M
This research model fine-tunes a roughly 600M Nemotron Conformer with hybrid RNNT and CTC decoding on about 200 hours of Bhili speech.
Restricted access
Streaming settings share one trained checkpoint
The model supports cache-aware streaming and offline inference. Its output layers must be restricted to the Bhili head. The card specifies mono 16 kHz audio and cautions about performance outside agricultural content.Bhili ASR Nemotron 600M official model card
Compare latency and transcript revisions together
Practical guidance. A shorter streaming delay can change how much context is available for a word. Measure both user-perceived delay and the final transcript on the same recordings.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Restricted access |
| Access | Downloadable |
| Model identifiers | bhili-asr-nemotron-600m |
| Access | Case-by-case approval for an unreleased research model; MIT metadata.Bhili ASR Nemotron 600M official model card |
| Streaming | Seven attention-context settings were trained together; latency can be selected during inference.Bhili ASR Nemotron 600M official model card |
| Language head | target_lang=bhb is required on every call.Bhili ASR Nemotron 600M official model card |
Related models
- Bhili ASR Canary 1.2B
AI4Bharat · Bhili Models
Research speech recognizer adapting a 1.2-billion-parameter Canary architecture to Dehvali Bhili.
Launch date not confirmedRestricted accessDownloadable - Bhili ASR Conformer 600M
AI4Bharat · Bhili Models
600-million-parameter IndicConformer adaptation trained on about 200 hours of Bhili speech.
Launch date not confirmedDownloadable