wave.readframes()与librosa.load()读取音频数据的差异原因咨询

Great question! The difference you're seeing comes down to how each library processes and represents audio data—let's break it down clearly:

1. Automatic Normalization (The Core Reason) #

Librosa's load() function automatically normalizes audio data to the [-1.0, 1.0] range by default. This is standard for audio processing libraries because floating-point values in this range are far easier to work with for operations like STFT, feature extraction, or neural network inputs.

Your WAV file uses 16-bit PCM encoding (the most common format), which stores raw sample values as integers between -32768 and 32767. Librosa takes these raw integers and divides each by 32768 to scale them into the [-1, 1] float range.

Looking at your output, you can see this exact relationship:

• Wave's first value is -1 → -1 / 32768 = -3.0517578e-05 (matches Librosa's first value)
• Wave's second value is 13 → 13 / 32768 ≈ 3.9672852e-04 (matches Librosa's second value)

2. Data Type Differences #

• wave.readframes() returns raw byte data, which you converted to np.int16—this is the original integer format the WAV file stores samples in.
• librosa.load() returns float32 values by default, even if you disable normalization (using normalize=False). Floating-point arithmetic is standard for audio processing due to better precision during complex operations.

3. Extra Librosa Processing (No Impact on Your Sample Length) #

Librosa handles a few more things behind the scenes that didn't affect your sample count here, but are useful to know:

• It converts stereo audio to mono by default (use mono=False to preserve channels). Your length matches wave's output, so your file is already mono.
• It resamples the audio to your specified sr=44100 if the original sample rate differs. Since your length matches, your file's native sample rate is already 44100.

Quick Verification Code #

To confirm this relationship, run this snippet—it scales Librosa's data back to the original integer range and checks if it matches the wave output:

import numpy as np
import librosa
import wave

sample_wave = './data/mywave.wav'

# Load with Librosa
a, sr = librosa.load(sample_wave, sr=44100)
# Load with wave
wav = wave.open(sample_wave)
data = wav.readframes(wav.getnframes())
b = np.frombuffer(data, dtype=np.int16)

# Scale Librosa's float data back to int16
a_scaled = (a * 32768).astype(np.int16)

# Check if they match
print(np.array_equal(a_scaled, b))  # Should print True!

内容的提问来源于stack exchange，提问作者whitebear