feat: use torchaudio instead of soundfile

docs/source/pretrained/whisper.py

@@ -1,3 +1,4 @@
+import einops
 from transformers import AutomaticSpeechRecognitionPipeline, pipeline
 
 from todd.utils import get_audio
@@ -6,6 +7,8 @@
     'https://github.com/SWivid/F5-TTS/raw/refs/heads/main/'
     'src/f5_tts/infer/examples/basic/basic_ref_zh.wav',
 )
+audio_array = audio.numpy()
+audio_array = einops.rearrange(audio_array, '1 t -> t')
 
 pipe: AutomaticSpeechRecognitionPipeline = pipeline(
     'automatic-speech-recognition',
@@ -14,17 +17,20 @@
     device_map='auto',
 )
 
-result = pipe(audio)
+result = pipe(audio_array)
 print(result)
 
-result = pipe(audio, generate_kwargs=dict(language='zh'))
+result = pipe(audio_array, generate_kwargs=dict(language='zh'))
 print(result)
 
-result = pipe(audio, generate_kwargs=dict(task='translate', language='en'))
+result = pipe(
+    audio_array,
+    generate_kwargs=dict(task='translate', language='en'),
+)
 print(result)
 
-result = pipe(audio, return_timestamps=True)
+result = pipe(audio_array, return_timestamps=True)
 print(result)
 
-result = pipe(audio, return_timestamps='word')
+result = pipe(audio_array, return_timestamps='word')
 print(result)

pyproject.toml

@@ -22,7 +22,6 @@ dependencies = [
     'opencv-python',
     'python-pptx',
     'pycocotools',
-    'soundfile',
     'tensorboard',
     'timm',
     'toml',
@@ -163,7 +162,6 @@ module = [
     'pptx.*',
     'scipy.*',
     'setuptools.*',
-    'soundfile.*',
     'torchvision.*',
     'transformers.*',
     'yapf.*',

todd/utils/networks.py

@@ -9,7 +9,8 @@
 import numpy as np
 import numpy.typing as npt
 import requests
-import soundfile as sf
+import torch
+import torchaudio
 from PIL import Image
 
 
@@ -25,5 +26,5 @@ def get_image(url: str) -> npt.NDArray[np.uint8]:
         return np.array(image)
 
 
-def get_audio(url: str) -> tuple[npt.NDArray[np.float64], int]:
-    return sf.read(get_bytes(url))
+def get_audio(url: str) -> tuple[torch.Tensor, int]:
+    return torchaudio.load(get_bytes(url))