feat(gallery): add experimental Fish Drama 3 (#12789)

Offer the community GGUF with Q4 experts and a Q8 backbone through
existing audio-cpp routing. Document cloning inputs, memory needs, and
the Fish Audio Community License.

Keep automatic task selection for voice cloning. The existing audio.cpp
pin defaults to top_k 4097, so no sampling override or backend bump is
needed.

Assisted-by: Codex:gpt-5

Co-authored-by: Ettore Di Giacinto <mudler@localai.io>

gallery/index.yaml

@@ -68908,6 +68908,52 @@
     - filename: audio-cpp/yue2-3b/sidecars/yue2-vae-config.json
       sha256: f0191bb9694009956de44e0c361a6f1334760be4c8f848e599bde242a54a0970
       uri: huggingface://audio-cpp/Yue2-3B-GGUF/sidecars/yue2-vae-config.json
+- name: audio-cpp-fish-drama3
+  url: github:mudler/LocalAI/gallery/virtual.yaml@master
+  urls:
+    - https://huggingface.co/fishaudio/drama-3
+    - https://huggingface.co/Jmica/Drama-3-GGUF
+    - https://huggingface.co/Jmica/Drama-3-GGUF/blob/17a5b40c5dc2c2be986f5371a7e650108b828e29/LICENSE
+    - https://github.com/0xShug0/audio.cpp
+  description: |
+    Fish Audio Drama 3 is an experimental community port served by audio-cpp.
+    This GGUF uses Q4_K_S experts and an 8-bit (Q8) backbone. The download is
+    approximately 24.6 GB; runtime memory requirements exceed the weight size.
+    Allow additional memory for inference and reference audio processing.
+
+    For voice cloning, supply a server-local reference WAV path in the OpenAI
+    `voice` field and its transcript in `params.reference_text` (or
+    `params.ref_text`). Cloning requires reference audio and a matching
+    transcript; plain TTS can run without a reference. No named preset voices
+    are provided. Use only reference audio for which you have rights and consent.
+
+    The Fish Audio Community License governs the weights, not MIT. Review its
+    use and redistribution terms before installation. This community port is
+    experimental and is not an official Fish Audio serving implementation.
+  license: other
+  tags:
+    - audio-cpp
+    - fish-drama3
+    - experimental
+    - tts
+    - text-to-speech
+    - voice-cloning
+    - gguf
+    - quantized
+  overrides:
+    backend: audio-cpp
+    known_usecases:
+      - tts
+    name: audio-cpp-fish-drama3
+    options:
+      - family:fish_drama3
+      - backend:best
+    parameters:
+      model: audio-cpp/drama-3-backbone-8bit-experts-Q4_K_S.gguf
+  files:
+    - filename: audio-cpp/drama-3-backbone-8bit-experts-Q4_K_S.gguf
+      sha256: 05af716baf78e7c20863b9ccebb3ebdeca9607dd3ea7d0b72c35079e675b212b
+      uri: https://huggingface.co/Jmica/Drama-3-GGUF/resolve/17a5b40c5dc2c2be986f5371a7e650108b828e29/drama-3-backbone-8bit-experts-Q4_K_S.gguf
 - name: audio-cpp-indextts-2.5
   url: github:mudler/LocalAI/gallery/virtual.yaml@master
   urls:

tests/test_fish_drama3_gallery.py

@@ -0,0 +1,49 @@
+# SPDX-License-Identifier: MIT
+"""Offline Drama 3 gallery regression. Requires PyYAML; downloads no weights."""
+import pathlib
+import unittest
+
+import yaml
+
+
+ROOT = pathlib.Path(__file__).resolve().parents[1]
+NAME = "audio-cpp-fish-drama3"
+FILE = "drama-3-backbone-8bit-experts-Q4_K_S.gguf"
+REVISION = "17a5b40c5dc2c2be986f5371a7e650108b828e29"
+SHA256 = "05af716baf78e7c20863b9ccebb3ebdeca9607dd3ea7d0b72c35079e675b212b"
+
+
+class Drama3GalleryTest(unittest.TestCase):
+    def test_gallery_entry(self):
+        entries = yaml.safe_load((ROOT / "gallery/index.yaml").read_text())
+        matches = [entry for entry in entries if entry["name"] == NAME]
+        self.assertEqual(len(matches), 1, "expected exactly one Drama 3 entry")
+        entry = matches[0]
+        self.assertEqual(entry["url"], "github:mudler/LocalAI/gallery/virtual.yaml@master")
+        yaml.safe_load((ROOT / "gallery/virtual.yaml").read_text())
+        self.assertEqual(entry["license"], "other")
+        self.assertIn("https://huggingface.co/Jmica/Drama-3-GGUF/blob/" + REVISION + "/LICENSE", entry["urls"])
+        self.assertIn("plain TTS can run without a reference", entry["description"])
+        self.assertTrue({"tts", "voice-cloning", "experimental", "gguf"} <= set(entry["tags"]))
+        overrides = entry["overrides"]
+        self.assertEqual(overrides["name"], NAME)
+        self.assertEqual(overrides["backend"], "audio-cpp")
+        self.assertEqual(overrides["known_usecases"], ["tts"])
+        # Do not pin task:tts: the router must still select voice cloning.
+        self.assertEqual(set(overrides["options"]), {"family:fish_drama3", "backend:best"})
+        # LLM top_k is not forwarded by the TTS RPC. The existing audio.cpp
+        # c6abb443 pin defaults to 4097 in fish_drama3/session.cpp:138.
+        self.assertEqual(overrides["parameters"], {"model": "audio-cpp/" + FILE})
+        self.assertNotIn("voices", overrides.get("tts", {}))
+        self.assertEqual(entry["files"], [{
+            "filename": "audio-cpp/" + FILE,
+            "sha256": SHA256,
+            "uri": "https://huggingface.co/Jmica/Drama-3-GGUF/resolve/" + REVISION + "/" + FILE,
+        }])
+        for text in ("experimental", "community port", "reference", "transcript",
+                     "24.6 GB", "memory", "Fish Audio Community License"):
+            self.assertIn(text, entry["description"])
+
+
+if __name__ == "__main__":
+    unittest.main()