NULLXES KLAREN 4E
Consolidated multimodal engineering checkpoint.
Contains:
- Language model
- Vision encoder and trained vision projector
- Whisper audio encoder and experimentally aligned audio projector
The audio projector was trained on a small English speech dataset. General speech recognition and multilingual audio understanding have not been established.
Custom model code is provided for loading and runtime integration. This is an experimental checkpoint, not a validated production ASR system.
Audio ASR V1.1
Includes a pretrained Whisper-small speech decoder and output projection alongside the previously consolidated language, vision and audio components.
Validation: WER 7.50%, CER 2.49% on 10 English recordings. This small validation set does not establish broad speech recognition performance.
The experimental audio projector remains present. ASR transcription uses the Whisper encoder-decoder path, not the experimental projector.
- Downloads last month
- -