Instinct Speech — Current Models
Collection
Current speech releases, with production packages first and earlier public research preserved. • 2 items • Updated
Instinct STT 0.1
Instinct STT 0.1 is an archived multilingual speech-recognition research checkpoint for Russian, Uzbek, and Kazakh. It is preserved for reproducibility and comparison; Instinct STT 1 is the current model family.
Checkpoint
- Full training-state checkpoint: step-265918.ckpt
- Global step: 265,918
- Size: 7,741,181,335 bytes
- SHA256: a6c8e9258589deba8e3ade406c94054c4daa05a670bea1414e23d7d109bd6699
- Languages: Russian, Uzbek, Kazakh
- Precision used during training: BF16
This artifact includes optimizer and scheduler state and is intended for controlled research restoration. It is not packaged as a drop-in production inference model.
Final evaluation snapshot
| Language | WER | CER |
|---|---|---|
| Russian | 20.61% | 6.15% |
| Uzbek | 49.52% | 16.00% |
| Kazakh | 41.90% | 13.40% |
These results belong to the original research protocol. Some evaluation material overlapped training, so they must not be presented as independent generalization benchmarks.
Recommended model
For the current multilingual ASR package, inference code, and staged checkpoints, use Instinct STT 1.
Access
This repository is public with manually approved access.