| base_model: | |
| - openai/whisper-small | |
| - Qwen/Qwen2.5-1.5B-Instruct | |
| library_name: pytorch | |
| pipeline_tag: feature-extraction | |
| license: apache-2.0 | |
| tags: | |
| - audio | |
| - speech-enhancement | |
| - multimodal | |
| - projector | |
| # Audio-Codec-LLM native audio projector (v3 smoke) | |
| `audio_projector.pt` is the trainable Whisper-to-Qwen continuous prefix projector | |
| from the v3 RTX 4090 smoke run. Whisper-small and Qwen2.5-1.5B-Instruct were loaded | |
| as real frozen base models; two optimizer steps reduced the smoke loss from | |
| 3.31035 to 2.70892 (mean 3.00964), with 3,859.58 MiB peak VRAM. | |
| This artifact proves the tensor path, backward pass, optimizer and serialization. | |
| Two steps are not a convergence experiment and do not establish enhancement | |
| quality. See `run_manifest.json` and the | |
| [source repository](https://github.com/Jatshi/llm-guided-speech-enhancement). | |