Use soundfile instead of torchaudio.

examples/generate.py

@@ -12,8 +12,7 @@
 
 from vocos_mlx import Vocos
 
-import torch
-import torchaudio
+import soundfile as sf
 
 SAMPLE_RATE = 24_000
 HOP_LENGTH = 256
@@ -37,9 +36,9 @@ def generate(
     f5tts = CFM.from_pretrained(model_name, vocab)
 
     # load reference audio
-    audio, sr = torchaudio.load(Path(ref_audio_path))
-    audio = mx.array(audio.numpy())
-    ref_audio_duration = audio.shape[1] / SAMPLE_RATE
+    audio, sr = sf.read(ref_audio_path)
+    audio = mx.array(audio)
+    ref_audio_duration = audio.shape[0] / SAMPLE_RATE
 
     rms = mx.sqrt(mx.mean(mx.square(audio)))
     if rms < TARGET_RMS:
@@ -55,7 +54,7 @@ def generate(
     vocos = Vocos.from_pretrained("lucasnewman/vocos-mel-24khz")
 
     wave, _ = f5tts.sample(
-        audio,
+        mx.expand_dims(audio, axis=0),
         text=text,
         duration=frame_duration,
         steps=32,
@@ -66,13 +65,13 @@ def generate(
     )
 
     # trim the reference audio
-    wave = wave[audio.shape[1] :]
-    generated_duration = len(wave) / SAMPLE_RATE
+    wave = wave[audio.shape[0]:]
+    generated_duration = wave.shape[0] / SAMPLE_RATE
     elapsed_time = datetime.datetime.now() - start_date
 
     print(f"Generated {generated_duration:.2f} seconds of audio in {elapsed_time}.")
 
-    torchaudio.save(output_path, torch.Tensor(np.array(wave)).unsqueeze(0), SAMPLE_RATE)
+    sf.write(output_path, np.array(wave), SAMPLE_RATE)
 
 
 if __name__ == "__main__":

pyproject.toml

@@ -7,15 +7,14 @@ requires = [
     "mlx",
     "numpy",
     "setuptools",
-    "torch",
-    "torchaudio",
+    "soundfile",
     "vocos-mlx"
 ]
 build-backend = "setuptools.build_meta"
 
 [project]
 name = "f5-tts-mlx"
-version = "0.0.2"
+version = "0.0.3"
 authors = [{name = "Lucas Newman", email = "lucasnewman@me.com"}]
 license = {text = "MIT"}
 description = "F5-TTS - MLX"