stretch-audio

High-quality time-stretching of audio files — make them faster or slower without changing pitch and without artifacts.

Reverse-engineered from Adobe Audition 2020's Izotope Omega engine (dvaaudiodsp.framework, dsp::stretch::omega namespace) and implemented as a pure Python/NumPy library.

Install

pip install stretch-audio

Usage

import numpy as np
from stretch_audio import time_stretch, time_stretch_multichannel

# Mono — slow down by 1.5×
output = time_stretch(audio, rate=1.5, sample_rate=44100)

# Mono — speed up by 2×
output = time_stretch(audio, rate=0.5, sample_rate=44100)

# Stereo — shape (2, N)
output = time_stretch_multichannel(stereo, rate=1.25, sample_rate=44100)

audio is a 1-D float32 or float64 NumPy array. The output is float32 with length round(len(audio) * rate).

Parameters

Parameter	Default	Description
rate	—	Stretch ratio. >1 = slower, <1 = faster. Clamped to [1/16, 16] in the engine.
sample_rate	44100	Audio sample rate in Hz
n_fft	auto	FFT window size (power of two). Auto-selected from sample_rate to target ~46 ms (2048 at 44100 Hz). Pass explicitly to override.
hop_length	n_fft // 4	Analysis hop — 75% overlap, matching the engine's hardcoded hop_divisor = 4
transient_sensitivity	1.0	0 = pure phase vocoder, higher = more transient locking. Engine modes: Music = 1.0, Speech = 0.5, Off = 0.0

Algorithm

The implementation matches the Transient-Aware Phase Vocoder (TAPV) used in Audition's Omega engine:

1. STFT analysis — periodic Hanning window, 75% overlap (PrecomputeWindows)
2. Transient detection — positive spectral flux on the power spectrum with an adaptive local-mean threshold and non-maximum suppression (TransientsInfo::DetectTransientsOfPowerSpectrum + DiscardWeakerBlocks)
3. Phase vocoder — for tonal frames, instantaneous frequency is estimated per bin and the phase accumulator is advanced by true_freq × synthesis_hop. For transient frames the accumulator is reset to the raw analysis phase, preserving attack shape exactly. (OmegaEngineImpl::ScheduleGranules)
4. Overlap-add synthesis — IFFT + pre-normalised synthesis window + OLA (DispatchOLA, PasteCurrentGranule)
5. Polyphase resampling — exact output length via sinc resampling (OmegaEngineImpl::ResampleOutput)

Auto window sizing

The engine internally precomputes 32 log-spaced window sizes from ~1 ms to ~100 ms and selects adaptively per frame. The fixed-window PV here targets ~46 ms — wide enough for good frequency resolution, narrow enough to avoid temporal smearing. n_fft scales with sample_rate:

Sample rate	Default n_fft	Window length
22050 Hz	1024	46 ms
44100 Hz	2048	46 ms
48000 Hz	2048	43 ms
96000 Hz	4096	43 ms

Development

git clone https://github.com/your-username/stretch-audio
cd stretch-audio
uv sync --dev
uv run pytest