如何用Python采样PC音频设备并检测特定输出音频？

Hey there! Let's tackle your two problems step by step, with practical code examples you can adapt for your use case.

1. Sampling Your PC's Audio Output Stream #

You mentioned struggling with sounddevice—the key here is targeting the loopback/立体声混音 device that captures audio being sent to your speakers, rather than a microphone input. Here's how to make it work:

First, install the required packages:

pip install sounddevice soundfile librosa numpy

Step 1: Find Your Loopback Device #

Run this script to list all audio devices and spot your system's loopback device (it’ll usually have "Loopback", "Stereo Mix", or "Monitor" in the name):

import sounddevice as sd

print("Available audio devices:")
for i, device in enumerate(sd.query_devices()):
    print(f"ID {i}: {device['name']}")
    # Check for loopback/monitor device (naming varies by OS)
    if "loopback" in device['name'].lower() or "stereo mix" in device['name'].lower():
        print(f"→ Found loopback device: ID {i}")

Step 2: Capture Audio in Real-Time #

Once you have the loopback device ID, use this code to capture and process audio in chunks (we’ll reuse this for sound detection later):

import sounddevice as sd
import numpy as np
import sys

# Replace with your loopback device ID from the previous step
LOOPBACK_DEVICE_ID = 2
SAMPLING_RATE = 44100
CHUNK_DURATION = 0.5  # Process audio in 0.5-second chunks
CHUNK_SIZE = int(SAMPLING_RATE * CHUNK_DURATION)

def audio_callback(indata, frames, time, status):
    if status:
        print(status, file=sys.stderr)
    # Convert stereo audio to mono for simpler processing
    mono_audio = np.mean(indata, axis=1)
    # Pass the chunk to your detection function
    detect_target_sound(mono_audio)

# Start the audio stream
with sd.InputStream(device=LOOPBACK_DEVICE_ID,
                    samplerate=SAMPLING_RATE,
                    blocksize=CHUNK_SIZE,
                    callback=audio_callback):
    print("Listening to audio output... Press Enter to stop.")
    input()

2. Detecting Short Target Sounds (1-2 Seconds) #

Chromaprint’s fpcalc struggles with short audio because it relies on longer temporal patterns for fingerprinting. Instead, we’ll use MFCC (Mel-Frequency Cepstral Coefficients)—a feature that captures the unique timbral characteristics of sound—paired with Dynamic Time Warping (DTW) to handle slight timing differences between the target and captured audio.

Step 1: Preprocess Your Target Audio #

First, extract MFCC features from your target sound file:

import librosa

TARGET_AUDIO_PATH = "target_sound.wav"
TARGET_SAMPLING_RATE = 44100

# Load target audio and convert to mono
target_audio, _ = librosa.load(TARGET_AUDIO_PATH, sr=TARGET_SAMPLING_RATE, mono=True)
# Normalize volume to reduce volume-based errors
target_audio = librosa.util.normalize(target_audio)

# Extract MFCC features + delta/delta-delta for temporal context
target_mfcc = librosa.feature.mfcc(y=target_audio, sr=TARGET_SAMPLING_RATE, n_mfcc=13)
target_mfcc_delta = librosa.feature.delta(target_mfcc)
target_mfcc_delta2 = librosa.feature.delta(target_mfcc, order=2)
# Combine all features into one array
target_features = np.concatenate([target_mfcc, target_mfcc_delta, target_mfcc_delta2], axis=0)

Step 2: Real-Time Detection with DTW #

Add this function to your stream code to compare captured audio chunks with the target features:

from librosa.sequence import dtw

def detect_target_sound(captured_audio):
    # Normalize captured audio
    captured_audio = librosa.util.normalize(captured_audio)
    
    # Extract features from the captured chunk
    captured_mfcc = librosa.feature.mfcc(y=captured_audio, sr=SAMPLING_RATE, n_mfcc=13)
    captured_mfcc_delta = librosa.feature.delta(captured_mfcc)
    captured_mfcc_delta2 = librosa.feature.delta(captured_mfcc, order=2)
    captured_features = np.concatenate([captured_mfcc, captured_mfcc_delta, captured_mfcc_delta2], axis=0)
    
    # Calculate DTW distance (lower value = more similar)
    try:
        dtw_distance, _, _, _ = dtw(captured_features.T, target_features.T)
        # Adjust this threshold based on your testing (lower = stricter match)
        THRESHOLD = 1000
        if dtw_distance < THRESHOLD:
            print("Target sound detected! Triggering action...")
            # Add your desired action here (e.g., simulate a keypress, run a script)
            # trigger_your_action()
    except Exception:
        # Skip chunks that are too short for valid comparison
        pass

Quick Optimization Tips #

• Tweak the Threshold: Test with different audio scenarios to find a threshold that balances false positives and negatives.
• Reduce Sampling Rate: If you need faster processing, lower the sampling rate (e.g., 22050 instead of 44100) to reduce feature computation time.
• Use Mel Spectrograms: For simpler, faster detection, replace MFCCs with a Mel spectrogram (remove the delta features and use librosa.feature.melspectrogram() instead).

内容的提问来源于stack exchange，提问作者RenegadeAndy