exploration of alternative wavtokenizer

Author: LostRuins

examples/tts/convert_pt_to_hf.py

@@ -154,7 +154,7 @@ def flatten_state_dict(state_dict, parent_key='', sep='.'):
     "architectures": [
         "WavTokenizerDec"
     ],
-    "hidden_size": 1282,
+    "hidden_size": 1282, # or 2402 for 40t/s
     "n_embd_features": 512,
     "n_ff": 2304,
     "vocab_size": 4096,

otherarch/tts_adapter.cpp

@@ -152,9 +152,9 @@ static std::vector<float> embd_to_audio(
         const int n_codes,
         const int n_embd,
         const int n_thread) {
-    const int n_fft = 1280;
-    const int n_hop = 320;
-    const int n_win = 1280;
+    const int n_hop = 600;
+    const int n_fft = n_hop*4; //its 1280 at 320, or 2400 at 600
+    const int n_win = n_hop*4;
     const int n_pad = (n_win - n_hop)/2;
     const int n_out = (n_codes - 1)*n_hop + n_win;