new model - Azure Cobalt (#4)

- new model - Azure Cobalt (e0d3ffd3485825d03fc722e2e608115ec86d9a0b)


Co-authored-by: Aster <[email protected]>

Azure.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:96460bc65045cdf0552c235872996f24c5adf3f83373cd594ea8970160bb384c
+size 409441960

Azure.py

@@ -0,0 +1,48 @@
+from fish_diffusion.datasets.hifisinger import HiFiSVCDataset
+from fish_diffusion.datasets.utils import get_datasets_from_subfolder
+
+_base_ = [
+    "./_base_/archs/hifi_svc.py",
+    "./_base_/trainers/base.py",
+    "./_base_/schedulers/exponential.py",
+    "./_base_/datasets/hifi_svc.py",
+]
+
+speaker_mapping = {
+    "Placeholder": 0,
+}
+
+model = dict(
+    type="HiFiSVC",
+    speaker_encoder=dict(
+        input_size=len(speaker_mapping),
+    ),
+)
+
+preprocessing = dict(
+    text_features_extractor=dict(
+        type="ContentVec",
+    ),
+    pitch_extractor=dict(
+        type="CrepePitchExtractor",
+        keep_zeros=False,
+        f0_min=40.0,
+        f0_max=2000.0,
+    ),
+    energy_extractor=dict(
+        type="RMSEnergyExtractor",
+    ),
+    augmentations=[
+        dict(
+            type="FixedPitchShifting",
+            key_shifts=[-5.0, 5.0],
+            probability=0.75,
+        ),
+    ],
+)
+
+trainer = dict(
+    # Disable gradient clipping, which is not supported by custom optimization
+    gradient_clip_val=None,
+    max_steps=1000000,
+)

config.yaml

@@ -83,4 +83,12 @@ models:
     readme: |
       This model is trained on a datset known as C and released under the [CC-BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/) license.
       It has a whispery, fluttery voice.
-    default_speaker: "c"
+    default_speaker: "c"
+
+  - name: "Azure Cobalt (Feminine)"
+    config: configs/Azure.py
+    checkpoint: checkpoints/Azure.ckpt
+    readme: |
+      This model is trained on a dataset known as Azure Cobalt and released under the [CC-BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/) license.
+      It has a whispery, fluttery voice.
+    default_speaker: "azure"