Update README.md

README.md

@@ -1,11 +1,14 @@
 # Spirit LM Inference Gradio Demo
 
-Copy the github repo, build the spiritlm python package and put models in `checkpoints` folder before running the script. I would suggest to use conda environment for this.
+Copy the github repo, build the [spiritlm](https://github.com/facebookresearch/spiritlm) python package and put models in `checkpoints` folder before running the script. I would suggest to use conda environment for this.
 
-You need around 15.5GB of VRAM to run the model with short output length and around 19GB to output 800 tokens.
+You need around 15.5GB of VRAM to run the model with 200 tokens output length and around 19GB to output 800 tokens.
 
-Edit: Audio to audio inference doesn't seem great. Potentially I am tokenizing the audio wrong. Could be also that model doesn't work well with audio IN audio OUT.
+If you're concerned about pickles from unknown uploader - grab them from a repo maintained by HF staffer - [https://huggingface.co/spirit-lm/Meta-spirit-lm](https://huggingface.co/spirit-lm/Meta-spirit-lm)
 
+Audio to audio inference doesn't seem good at all. Potentially I am tokenizing the audio wrong. Could be also that model doesn't work well with audio IN audio OUT.
+
+Script here works with just single speaker - if you know how to get other speakers let me know and I'll update it.
 
 ```python
 import gradio as gr