-

app.py

@@ -50,6 +50,12 @@ def get_mytts(text: str, voice: str, base_url: str):
     )
     return response.json()
 
+if is_gpu:
+    @spaces.GPU
+    def tts_gpu(text: str, voice: str):
+        wav, sampling_rate = generate_tts(text, voice, 1)
+        audio_array = np.array(wav, dtype=np.float32).reshape(1, -1)
+        return (sampling_rate, audio_array)
 
 async def response(
     audio: tuple[int, np.ndarray],
@@ -92,9 +98,7 @@ async def response(
         audio_array = wav.reshape(1, -1)
         yield (result['sampling_rate'], audio_array)
     elif is_gpu and not tts_url:
-        wav, sampling_rate = generate_tts(text, voice, 1)
-        audio_array = np.array(wav, dtype=np.float32).reshape(1, -1)
-        yield (sampling_rate, audio_array)
+        yield tts_gpu(text, voice)
     else:
         raise ValueError("No TTS URL provided and GPU not available")
 

model.py

@@ -75,7 +75,6 @@ CLONE_VOICES = {
 
 f5tts = F5TTS(ckpt_file=model_path, vocab_file='vocab.txt', device='cuda')
 
-@spaces.GPU()
 def generate_tts(gen_text, voice, speed):
     wav, sampling_rate, _ = f5tts.infer(
         ref_file=f'ref/{CLONE_VOICES[voice]["path"]}',

requirements.txt

@@ -28,7 +28,7 @@ wandb
 x_transformers>=1.31.14
 matplotlib
 accelerate>=0.33.0,!=1.7.0
-bitsandbytes>0.37.0; platform_machine != 'arm64' and platform_system != 'Darwin'
+bitsandbytes>0.37.0
 click
 datasets
 ema_pytorch>=0.5.2