Add B3 FP16 speaker model: NOTICE
Browse files- redimnet2-b3-fp16-v1/NOTICE +14 -0
redimnet2-b3-fp16-v1/NOTICE
ADDED
|
@@ -0,0 +1,14 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
ReDimNet2-B3 mixed speaker embeddings
|
| 2 |
+
Copyright (c) 2026 Palabra.ai
|
| 3 |
+
Upstream: https://github.com/PalabraAI/redimnet2
|
| 4 |
+
License: MIT (included as LICENSE)
|
| 5 |
+
|
| 6 |
+
The waveform feature implementation in the upstream repository also carries:
|
| 7 |
+
Copyright (c) 2024 ID R&D, Inc.
|
| 8 |
+
That implementation is distributed under the same MIT permission and warranty
|
| 9 |
+
terms included in LICENSE. Its waveform preprocessing was exported unchanged
|
| 10 |
+
from the evaluated self-contained ONNX graph.
|
| 11 |
+
|
| 12 |
+
SpeechlessDo modifications: split FP32 waveform preprocessing from the original
|
| 13 |
+
network, convert the network to native CoreML MLProgram FP16 with fixed-length
|
| 14 |
+
features and a separate FP16-weight CPU path for short utterances, and preserve float32 model inputs/outputs. No additional training.
|