| This model is licensed under the Fish Audio Research License, Copyright © 39 AI, INC. All Rights Reserved. | |
| This repository is a Derivative Work, under the Fish Audio Research License, | |
| of fishaudio/s2-pro (https://huggingface.co/fishaudio/s2-pro). | |
| Changes made: | |
| - A LoRA adapter (r=32, alpha=16) on the attention and mlp modules of both the | |
| slow and fast transformers. | |
| - 26 atomic input-only tokens for Hebrew IPA symbols, each with its own | |
| trainable embedding, added to the tokenizer (155,774 -> 155,800 tokens) and | |
| trained alongside the LoRA weights. The output vocabulary is unchanged. | |
| - A fix to an embedding-scale mismatch between the training and inference | |
| paths in fish_speech/models/text2semantic/llama.py (upstream of this | |
| adapter's own training), which applies to any S2-Pro fine-tune. | |
| The full license text is at: | |
| https://github.com/maxmelichov/fish-speech/blob/main/LICENSE | |