Downloads · 30 days
27
21% of all-time downloads
octopusmegalopod/experimental-tuned-english-asr
experimental-tuned-english-asr is a machine learning model from octopusmegalopod. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
How far can we go with a Q80 1.7B model? Can an mmproj even handle Q80? Surprisingly, the answer is yes. At least for audio.
Downloads · 30 days
27
21% of all-time downloads
All-time downloads
128
Public
Repo size
2.2 GB
Likes
0
Public
Click a slice to open those files.
.gguf2.2 GB · 100%
From the Hugging Face model README
How far can we go with a Q8_0 1.7B model? Can an mmproj even handle Q8_0? Surprisingly, the answer is yes. At least for audio.
This is a fine-tuned Qwen3 1.7B English-only ASR model using some experimental tricks. The biggest one being to fully finetune the audio encoder but only using LoRA on the language model. Unsurprisingly, this leads to a plateau on certain language modeling aspects e.g: punctuation. So for CWER this model would likely suffer.
Also of note is that despite being LoRA, the language decoder IS still changing substantially - it's not just minor adaptions. For example, forcing the decoder to use language "english" (never seen in training data) instead of "English" can cause it to spiral into an infinite loop.
Host it in llama-cpp like so: "llama-server" -m "llm-qxy.gguf" --mmproj "mmproj-qxy.gguf" -c 6000 -ngl 99
Or just use this tool to quickly convert something into an SRT with it.