Phase 1: Audio Processing & Streaming · 45 min · Python · scipy · libsamplerate

Audio Resampling and Format Conversion

Resampling is where audio quality lives or dies. Get it wrong and your ASR accuracy drops 20%.

Real-world application: Resampling quality directly impacts ASR accuracy — a measurable engineering decision.

What you will learn

Resample audio between sample rates without introducing artifacts
Use scipy.signal.resample and libsamplerate for high-quality resampling
Build the μ-law 8kHz → PCM 16kHz/24kHz conversion pipeline
Handle edge cases: clipping, aliasing, DC offset

The Problem

Telephony audio arrives at 8kHz. ASR needs 16kHz. TTS outputs 24kHz. WebRTC wants 48kHz. Every transition requires resampling — and bad resampling introduces artifacts that degrade ASR accuracy and audio quality.

The most common resampling bug: aliasing. When downsampling without a low-pass filter, high frequencies fold back into the audible range as artifacts. The audio sounds "metallic" or "underwater."

The Concept

Why Resampling Is Needed

The Nyquist Theorem and Aliasing

The Nyquist theorem states: to represent a frequency f, you need a sample rate of at least 2f. If you downsample from 48kHz to 16kHz without filtering, frequencies above 8kHz (the new Nyquist) will alias — fold back into the 0-8kHz range as false frequencies.

You're downsampling from 48kHz to 8kHz. What happens to a 10kHz tone in the audio?

Aliasing: a 10kHz tone in 48kHz audio, when downsampled to 8kHz without filtering, folds around the new Nyquist frequency (4kHz for 8kHz sample rate). 10kHz → 8000 - (10000 - 8000) = 6000Hz. Actually, the Nyquist for 8kHz is 4kHz, so 10kHz folds to |10000 - 2*4000| = 2000Hz. The fix: apply a low-pass filter at 4kHz before downsampling to remove all frequencies above the new Nyquist.

Upsampling vs Downsampling

Direction	Risk	Solution
Upsampling (8kHz → 16kHz)	No aliasing (more room)	Interpolation (linear, sinc)
Downsampling (48kHz → 16kHz)	Aliasing	Low-pass filter + decimation

Resampling Methods

Method	Quality	Speed	Use Case
Linear interpolation	Fair	Fast	Quick prototyping
`scipy.signal.resample`	Good	Medium	General purpose (FFT-based)
`scipy.signal.resample_poly`	Good	Fast	Polyphase filtering
libsamplerate (secret rabbit)	Excellent	Fast	Production quality
SoX resampling	Excellent	Medium	Offline processing

Anti-Aliasing Filter

Before downsampling, apply a low-pass filter at the target Nyquist frequency:

from scipy.signal import resample_poly, butter, sosfilt

def resample_with_antialias(audio, from_rate, to_rate):
    if from_rate == to_rate:
        return audio

    if to_rate < from_rate:
        # Downsampling: apply anti-aliasing filter
        cutoff = to_rate / 2  # New Nyquist
        sos = butter(10, cutoff / (from_rate / 2), output='sos')
        audio = sosfilt(sos, audio.astype(np.float32))

    # Use polyphase resampling
    from scipy.signal import resample_poly
    gcd = np.gcd(from_rate, to_rate)
    up = to_rate // gcd
    down = from_rate // gcd
    return resample_poly(audio, up, down).astype(np.int16)

Unlock the full lesson

You've read the first 2 sections. The rest of this lesson covers Build, Check Yourself, Key Terms & Next — plus a hands-on lab, quiz, and project artifact.

Create a free account — the opening phases of 29 of 35 courses are free, no credit card.