Upload folder using huggingface_hub

app.py

@@ -10,6 +10,8 @@ from matcha.models.matcha_tts import MatchaTTS
 from matcha.hifigan.models import Generator as HiFiGAN
 from matcha.hifigan.config import v1
 from matcha.hifigan.env import AttrDict
+from matcha.text import text_to_sequence
+from matcha.utils.utils import intersperse
 
 HF_TOKEN = os.getenv("HF_TOKEN")
 
@@ -42,11 +44,38 @@ model, vocoder = load_models()
 
 @torch.inference_mode()
 def process(text):
-    # Basic normalization for Kashmiri script
+    # 1. Kashmiri script normalization
     text = text.replace("ي", "ی").replace("ك", "ک").strip()
-    output = model.synthesise(text, n_timesteps=10, temperature=0.667, length_scale=1.0)
+    
+    # 2. Convert text to sequence using the correct cleaner
+    # We use 'basic_cleaners' here because the model was trained to 
+    # map Kashmiri characters directly to audio features.
+    cleaner = "basic_cleaners" 
+    x = torch.tensor(
+        intersperse(text_to_sequence(text, [cleaner])[0], 0),
+        dtype=torch.long,
+        device=DEVICE,
+    )[None]
+    x_lengths = torch.tensor([x.shape[-1]], dtype=torch.long, device=DEVICE)
+    
+    # 3. Generate Mel-spectrogram
+    output = model.synthesise(
+        x, 
+        x_lengths, 
+        n_timesteps=10, 
+        temperature=0.667, 
+        length_scale=1.0
+    )
+    
+    # 4. Generate Audio Waveform
     audio = vocoder(output['mel']).clamp(-1, 1).cpu().squeeze().numpy()
-    sf.write("out.wav", audio, 22050)
-    return "out.wav"
+    output_path = "out.wav"
+    sf.write(output_path, audio, 22050)
+    return output_path
 
-gr.Interface(fn=process, inputs="text", outputs="audio", title="Kashmiri TTS").launch()
+gr.Interface(
+    fn=process, 
+    inputs=gr.Textbox(label="Kashmiri Text"), 
+    outputs=gr.Audio(label="Audio", type="filepath"),
+    title="GAASH-Lab: Kashmiri TTS"
+).launch()