These Fish Audio Materials are licensed under the Fish Audio Community License, Version 1.0. Copyright © 39 AI, INC. All Rights Reserved.

Built with Fish Audio.

The attribution above applies to the Fish Audio Materials: Fish Audio Drama 3 and its audio codec, from
https://huggingface.co/fishaudio/drama-3 at revision e0a9218aa48888270f06ef1ed9ea67622c0e17d2, of which the files in
this repository are a Derivative Work. A copy of the Fish Audio Community License, Version 1.0, is in LICENSE.

The files in this repository were made by sakasegawa for speech.cpp (https://github.com/nyosegawa/speech.cpp), which
changed the Fish Audio Materials as follows:

- The model's weights, released with the routed experts in FP8 (blocks of 128 x 128 values with a float32 scale each)
  and the rest in BF16, and the audio codec, released mostly in float32, were converted to one GGUF file in float32, in
  speech.cpp's own layout: each 128 x 128 block of the experts multiplied by its scale, each layer's experts of a
  projection stacked into one tensor, the codec's weight normalization folded into its weights, some matrices split
  or reordered, and the output head and the audio decoder's code embeddings cut to the rows the model reads.
- Each file in this repository was quantized from that float32 file by speech.cpp's `speech quantize` to the weight
  type in its name (F16, Q8_0, Q6_K, Q5_K or Q4_K), which changes the values of the weights it quantizes.