update

main.py

@@ -22,6 +22,7 @@ import shutil
 import tempfile
 import time
 from typing import Dict, Tuple
+import uuid
 import zipfile
 
 import gradio as gr
@@ -30,11 +31,11 @@ import librosa
 import librosa.display
 import matplotlib.pyplot as plt
 import numpy as np
+from scipy.io import wavfile
 
 import log
 from project_settings import environment, project_path, log_directory
 from toolbox.os.command import Command
-from toolbox.torchaudio.models.dfnet.inference_dfnet import InferenceDfNet
 from toolbox.torchaudio.models.dfnet2.inference_dfnet2 import InferenceDfNet2
 from toolbox.torchaudio.models.dtln.inference_dtln import InferenceDTLN
 from toolbox.torchaudio.models.frcrn.inference_frcrn import InferenceFRCRN
@@ -79,6 +80,28 @@ def get_args():
     return args
 
 
+def save_input_audio(sample_rate: int, signal: np.ndarray) -> str:
+    if signal.dtype != np.int16:
+        raise AssertionError(f"only support dtype np.int16, however: {signal.dtype}")
+    temp_audio_dir = Path(tempfile.gettempdir()) / "input_audio"
+    temp_audio_dir.mkdir(parents=True, exist_ok=True)
+    filename = temp_audio_dir / f"{uuid.uuid4()}.wav"
+    filename = filename.as_posix()
+    wavfile.write(
+        filename,
+        sample_rate, signal
+    )
+    return filename
+
+
+def convert_sample_rate(signal: np.ndarray, sample_rate: int, target_sample_rate: int):
+    filename = save_input_audio(sample_rate, signal)
+
+    signal, _ = librosa.load(filename, sr=target_sample_rate)
+    signal = np.array(signal * (1 << 15), dtype=np.int16)
+    return signal
+
+
 def shell(cmd: str):
     return Command.popen(cmd)
 
@@ -131,6 +154,10 @@ def when_click_denoise_button(noisy_audio_file_t = None, noisy_audio_microphone_
     noisy_audio_t: Tuple = noisy_audio_file_t or noisy_audio_microphone_t
 
     sample_rate, signal = noisy_audio_t
+    if sample_rate != 8000:
+        signal = convert_sample_rate(signal, sample_rate, 8000)
+        sample_rate = 8000
+
     audio_duration = signal.shape[-1] // 8000
 
     # Test: 使用 microphone 时，显示采样率是 44100，但 signal 实际是按 8000 的采样率的。