在腾讯云云服务器CVM上玩转rnnoise

原创

buzzfrog

修改于 2025-04-26 15:46:46

18600

代码可运行

文章被收录于专栏：云上修行云上修行

运行总次数：0

代码可运行

rnnoise介绍

RNNoise是一种用于语音信号降噪的开源库，它使用递归神经网络（RNN）来实现噪声抑制。由Jean-Marc Valin开发，RNNoise结合了传统数字信号处理（DSP）和深度学习技术，旨在实时去除语音通信中的背景噪声。

为什么推荐rnnoise

专业优化：
专为语音设计的RNN架构
包含预训练好的噪声模型
实时处理延迟仅5ms
性能优势：
计算量比自行实现低10倍以上
已针对CPU指令集优化（SSE/AVX）
内存占用小于2MB
实用特性：
自动处理采样率转换
支持帧处理模式（适合实时流）
内置语音活动检测（VAD）

购买腾讯云云服务器

依赖安装

通过本地ssh ubuntu@服务器IP地址，远程登录到腾讯云云服务器上。在服务器中，执行如下指令：

sudo apt update
sudo apt-get update
sudo apt-get install git
sudo apt-get install autoconf
sudo apt-get install libtool

安装rnnoise

git clone https://github.com/xiph/rnnoise
cd rnnoise
./autogen.sh
./configure
make
sudo make install

configure:
------------------------------------------------------------------------
  rnnoise 0.2-22-g70f1d25: Automatic configuration OK.

    Assertions ................... no

    Hidden visibility ............ yes

    API code examples ............ yes
    API documentation ............ yes
------------------------------------------------------------------------

make会有很多warning，这里仅演示流程，可以忽略。如果需要提升性能，可以使用：

make CFLAGS="-march=native"

python3 wrapper程序

由于rnnoise编译出来是.so库，这里给一个python3的wrapper程序供python工程师使用。

首先需要安装依赖库

pip3 install numpy
pip3 install scipy

下面是pydenoise.py的源代码。

import numpy as np
import ctypes
from scipy.signal import resample


class RNNoiseWrapper:
    def __init__(self, lib_path):
        self.rnnoise_lib = ctypes.CDLL(lib_path)
        # RNNOISE_EXPORT DenoiseState *rnnoise_create(RNNModel *model);
        self.rnnoise_lib.rnnoise_create.argtypes = [ctypes.c_void_p]
        self.rnnoise_lib.rnnoise_create.restype = ctypes.c_void_p

        # RNNOISE_EXPORT float rnnoise_process_frame(DenoiseState *st, float *out, const float *in);
        self.rnnoise_lib.rnnoise_process_frame.argtypes = [
            ctypes.c_void_p,
            ctypes.POINTER(ctypes.c_float), 
            ctypes.POINTER(ctypes.c_float)
        ]
        self.rnnoise_lib.rnnoise_process_frame.restype = ctypes.c_float

        # RNNOISE_EXPORT void rnnoise_destroy(DenoiseState *st);
        self.rnnoise_lib.rnnoise_destroy.argtypes = [ctypes.c_void_p]

        self.state = self.rnnoise_lib.rnnoise_create(None)
    
    def detect_sample_rate_from_frame(self, frame):
        """Detect the sample rate based on the length of the frame."""
        frame_length = len(frame)
        if frame_length == 160:
            return 16000
        elif frame_length == 240:
            return 24000
        elif frame_length == 480:
            return 48000
        else:
            raise ValueError("Unsupported frame length. Expected 160, 240, or 480 for 10ms at 16000Hz, 24000Hz or 48000Hz.")

    def process_frame(self, input_frame):
        """ Process a frame of audio data with RNNoise, supporting multiple sample rates. """
        input_frame = np.asarray(input_frame, dtype=np.float32)
        
        # Detect the sample rate of the input frame
        input_sample_rate = self.detect_sample_rate_from_frame(input_frame)

        # Resample the input frame to 48000Hz if necessary
        original_length = len(input_frame)
        if input_sample_rate != 48000:
            input_frame = resample(input_frame, 480)
        
        # Allocate output buffer
        output_frame = np.zeros(480, dtype=np.float32)
        
        # Convert numpy array to ctypes pointers
        input_ptr = input_frame.ctypes.data_as(ctypes.POINTER(ctypes.c_float))
        output_ptr = output_frame.ctypes.data_as(ctypes.POINTER(ctypes.c_float))

        # Process the frame
        self.rnnoise_lib.rnnoise_process_frame(self.state, output_ptr, input_ptr)

        if input_sample_rate != 48000:
            output_frame = resample(output_frame, original_length)
        return output_frame

    def __del__(self):
        if self.state:
            self.rnnoise_lib.rnnoise_destroy(self.state)

# Usage Example
if __name__ == "__main__":
    lib_path = "/usr/local/lib/librnnoise.so"

    # Create RNNoise wrapper instance
    rnnoise_wrapper = RNNoiseWrapper(lib_path)

    # Create example input frames for different sample rates
    frame_16000Hz = np.random.randn(160).astype(np.float32)
    frame_24000Hz = np.random.randn(240).astype(np.float32)
    frame_48000Hz = np.random.randn(480).astype(np.float32)

    # Process frames
    processed_frame_16000 = rnnoise_wrapper.process_frame(frame_16000Hz)
    processed_frame_24000 = rnnoise_wrapper.process_frame(frame_24000Hz)
    processed_frame_48000 = rnnoise_wrapper.process_frame(frame_48000Hz)

    # Output the processed frames size
    print(f"Processed frame (16000Hz frame) length: {len(processed_frame_16000)}")
    print(f"Processed frame (24000Hz frame) length: {len(processed_frame_24000)}")
    print(f"Processed frame (48000Hz frame) length: {len(processed_frame_48000)}")