supports 24kHz

app.py

@@ -28,6 +28,15 @@ _ = freevc.eval()
 _ = utils.load_checkpoint("checkpoints/freevc.pth", freevc, None)
 smodel = SpeakerEncoder('speaker_encoder/ckpt/pretrained_bak_5805000.pt')
 
+print("Loading FreeVC(24k)...")
+hps = utils.get_hparams_from_file("configs/freevc-24.json")
+freevc_24 = SynthesizerTrn(
+    hps.data.filter_length // 2 + 1,
+    hps.train.segment_size // hps.data.hop_length,
+    **hps.model).to(device)
+_ = freevc_24.eval()
+_ = utils.load_checkpoint("checkpoints/freevc-24.pth", freevc_24, None)
+
 print("Loading FreeVC-s...")
 hps = utils.get_hparams_from_file("configs/freevc-s.json")
 freevc_s = SynthesizerTrn(
@@ -45,7 +54,7 @@ def convert(model, src, tgt):
         # tgt
         wav_tgt, _ = librosa.load(tgt, sr=hps.data.sampling_rate)
         wav_tgt, _ = librosa.effects.trim(wav_tgt, top_db=20)
-        if model == "FreeVC":
+        if model == "FreeVC" or model == "FreeVC (24kHz)":
             g_tgt = smodel.embed_utterance(wav_tgt)
             g_tgt = torch.from_numpy(g_tgt).unsqueeze(0).to(device)
         else:
@@ -67,23 +76,28 @@ def convert(model, src, tgt):
         # infer
         if model == "FreeVC":
             audio = freevc.infer(c, g=g_tgt)
-        else:
+        elif model == "FreeVC-s":
             audio = freevc_s.infer(c, mel=mel_tgt)
+        else:
+            audio = freevc_24.infer(c, g=g_tgt)
         audio = audio[0][0].data.cpu().float().numpy()
-        write("out.wav", hps.data.sampling_rate, audio)
+        if model == "FreeVC" or model == "FreeVC-s":
+            write("out.wav", hps.data.sampling_rate, audio)
+        else:
+            write("out.wav", 24000, audio)
     out = "out.wav"
     return out
     
-model = gr.Dropdown(choices=["FreeVC", "FreeVC-s"], value="FreeVC",type="value", label="Model") 
+model = gr.Dropdown(choices=["FreeVC", "FreeVC-s", "FreeVC (24kHz)"], value="FreeVC",type="value", label="Model") 
 audio1 = gr.inputs.Audio(label="Source Audio", type='filepath')
 audio2 = gr.inputs.Audio(label="Reference Audio", type='filepath')
 inputs = [model, audio1, audio2]
 outputs =  gr.outputs.Audio(label="Output Audio", type='filepath')
 
 title = "FreeVC"
-description = "Gradio Demo for FreeVC: Towards High-Quality Text-Free One-Shot Voice Conversion. To use it, simply upload your audio, or click the example to load. Read more at the links below. Note: It seems that the WavLM checkpoint in HuggingFace is a little different from the one used to train FreeVC, which may degrade the performance a bit. In addition, speaker similarity can be largely affected if there are too much silence in the reference audio, so please <strong>trim</strong> it before submitting."
+description = "Gradio Demo for FreeVC: Towards High-Quality Text-Free One-Shot Voice Conversion. To use it, simply upload your audio, or click the example to load. Read more at the links below. Note: It seems that the WavLM checkpoint in HuggingFace is a little different from the one used to train FreeVC, which may degrade the performance a bit."
 article = "<p style='text-align: center'><a href='https://arxiv.org/abs/2210.15418' target='_blank'>Paper</a> | <a href='https://github.com/OlaWod/FreeVC' target='_blank'>Github Repo</a></p>"
 
-examples=[["FreeVC", 'p225_001.wav', 'p226_002.wav'], ["FreeVC-s", 'p226_002.wav', 'p225_001.wav']]
+examples=[["FreeVC", 'p225_001.wav', 'p226_002.wav'], ["FreeVC-s", 'p226_002.wav', 'p225_001.wav'], ["FreeVC (24kHz)", 'p225_001.wav', 'p226_002.wav']]
 
 gr.Interface(convert, inputs, outputs, title=title, description=description, article=article, examples=examples, enable_queue=True).launch()

checkpoints/freevc-24.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2ff0ebf6bde90bf3f13518f49c204348d2f683ffb9fd31b24f59a2b302998862
+size 472644351

configs/freevc-24.json

@@ -0,0 +1,54 @@
+{
+  "train": {
+    "log_interval": 200,
+    "eval_interval": 10000,
+    "seed": 1234,
+    "epochs": 10000,
+    "learning_rate": 2e-4,
+    "betas": [0.8, 0.99],
+    "eps": 1e-9,
+    "batch_size": 64,
+    "fp16_run": false,
+    "lr_decay": 0.999875,
+    "segment_size": 8640,
+    "init_lr_ratio": 1,
+    "warmup_epochs": 0,
+    "c_mel": 45,
+    "c_kl": 1.0,
+    "use_sr": true,
+    "max_speclen": 128,
+    "port": "8008"
+  },
+  "data": {
+    "training_files":"filelists/train.txt",
+    "validation_files":"filelists/val.txt",
+    "max_wav_value": 32768.0,
+    "sampling_rate": 16000,
+    "filter_length": 1280,
+    "hop_length": 320,
+    "win_length": 1280,
+    "n_mel_channels": 80,
+    "mel_fmin": 0.0,
+    "mel_fmax": null
+  },
+  "model": {
+    "inter_channels": 192,
+    "hidden_channels": 192,
+    "filter_channels": 768,
+    "n_heads": 2,
+    "n_layers": 6,
+    "kernel_size": 3,
+    "p_dropout": 0.1,
+    "resblock": "1",
+    "resblock_kernel_sizes": [3,7,11],
+    "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
+    "upsample_rates": [10,6,4,2],
+    "upsample_initial_channel": 512,
+    "upsample_kernel_sizes": [16,16,4,4],
+    "n_layers_q": 3,
+    "use_spectral_norm": false,
+    "gin_channels": 256,
+    "ssl_dim": 1024,
+    "use_spk": true
+  }
+}