Changing the playback speed of audio without affecting its pitch is a fascinating problem in digital signal processing that requires sophisticated time-stretching algorithms. The naive approach to speed adjustment—simply playing back samples faster or slower—inherently changes both speed and pitch simultaneously, because pitch is a direct function of frequency, and frequency is determined by how many wave cycles occur per second. Playing audio at double speed doubles the frequency of every component, raising pitch by exactly one octave. This fundamental coupling between time and frequency is why early tape-based speed changes always produced the characteristic chipmunk (sped up) or slowed-down villain (slowed down) vocal effect.

Modern time-stretching algorithms break this coupling through a variety of techniques. The most widely used approach in real-time applications is the WSOLA (Waveform Similarity Overlap-Add) algorithm, which operates in the time domain by segmenting audio into small overlapping windows, then repositioning these windows in time while maintaining their original frequency content. To stretch audio (slow it down), WSOLA duplicates and overlaps segments; to compress audio (speed it up), it removes segments. The key innovation is the similarity-based alignment: rather than blindly splicing at fixed positions, the algorithm searches for optimal splice points where the waveform shapes match between adjacent segments, minimizing audible discontinuities and artifacts.

An alternative approach is the phase vocoder, which operates in the frequency domain using the Short-Time Fourier Transform (STFT). The audio is decomposed into overlapping spectral frames, each frame is phase-adjusted to maintain coherent frequency relationships at the modified time scale, and the frames are resynthesized using inverse FFT with modified overlap spacing. Phase vocoders excel at handling complex polyphonic audio and extreme speed ratios, but can introduce characteristic phasiness or metallic artifacts if phase coherence is not carefully maintained across frames.

The quality of time stretching depends heavily on the speed ratio and the audio content. Speech, with its relatively simple spectral content and clear transient structure, stretches extremely well at ratios from 0.5x to 2x. Music with complex harmonic content and prominent percussive elements presents a greater challenge, as transients can become smeared and harmonic relationships can develop subtle artifacts. Modern high-quality implementations often use hybrid approaches, combining time-domain processing for transient-rich segments with frequency-domain processing for sustained tonal content, producing transparent results across a wide range of speed modifications.