torch.hann_window
function hann_window(window_length: number, options?: WindowOptions): Tensorfunction hann_window(window_length: number, periodic: boolean, options?: WindowOptions): TensorHann (Hanning) window: symmetric window with zero endpoints for spectral analysis.
One of the most commonly used windows in signal processing. Applies smooth weighting that tapers to zero at the edges, reducing spectral leakage from strong signals. Essential for:
- Spectral analysis: Standard window for FFT analysis (minimal side lobes)
- Audio processing: Windowing before STFT for time-frequency analysis
- Signal detection: Finding weak signals near strong signals (low side lobe)
- Filter design: Designing FIR filters with good frequency response
- Vibration analysis: Detecting faults in mechanical systems
- Music analysis: Extracting frequency content while minimizing leakage
- Speech processing: Splitting speech into frames for analysis
Why Hann? Good balance between main lobe width (~4π/N) and side lobe level (-32 dB). Zeros at endpoints eliminate DC discontinuity when frames are repeated.
Periodic vs Non-periodic:
- periodic=True (default): For STFT/spectral analysis (seamless repetition)
- periodic=False: For filter design (symmetric, includes both endpoints)
- Symmetric: Window is symmetric around center point (zero is symmetric)
- Zero endpoints: Both first and last values are zero (reduces DC)
- Unity peak: Maximum value is 1.0 (at the center)
- Common choice: Default window for most spectral analysis applications
- Periodic variant: For STFT, use periodic=true (default) for seamless tiling
- Spectral leakage: Reduces but doesn't eliminate leakage from strong signals
- Not optimal for all tasks: Blackman has lower sidelobes (at cost of wider main lobe)
- Periodic vs symmetric: Different formulas; choose based on application
- Amplitude loss: Window reduces signal amplitude; must compensate in analysis
- Edge effects: Zeros at endpoints can hide important edge information
Parameters
window_lengthnumber- Size of the window (number of samples)
optionsWindowOptionsoptional- Window options: -
periodic: If true, computes periodic window for analysis (default: true) -dtype: Output data type (default: 'float32') -device: Where to place tensor (default: 'webgpu')
Returns
Tensor– 1D tensor of shape [window_length] with values in [0, 1]Examples
// Basic window: standard Hann for spectral analysis
const window = torch.hann_window(10);
// [0, 0.117, 0.413, 0.75, 0.970, 0.970, 0.75, 0.413, 0.117, 0]
// Note: zeros at endpoints minimize spectral leakage// STFT windowing: split signal into overlapping frames
const signal = torch.randn(16000); // 1 second at 16kHz
const hop_length = 512;
const window_size = 2048;
const window = torch.hann_window(window_size, { periodic: true });
// Extract frame and apply window
const frame = signal.slice(0, 0, window_size);
const windowed = frame.mul(window); // Element-wise multiply
// Windowed frame ready for FFT// Spectral analysis: visualize frequency content with Hann windowing
const audio = torch.randn(4096); // Audio sample
const window = torch.hann_window(4096);
const windowed = audio.mul(window);
const spectrum = torch.abs(torch.fft(windowed));
// Hann window reduces artificial spectral peaks from discontinuities// Filter design: create FIR filter with Hann window
const ideal_response = torch.sinc(torch.arange(-25, 26) / 5.0); // Ideal filter
const hann = torch.hann_window(51, { periodic: false });
const actual_filter = ideal_response.mul(hann);
// Window smooths sharp ideal response, reduces ringing// Comparing periodic vs non-periodic Hann windows
const periodic_hann = torch.hann_window(10, { periodic: true });
const symmetric_hann = torch.hann_window(10, { periodic: false });
// periodic: starts near 0, ends near 0 (seamless tiling for STFT)
// symmetric: exactly 0 at both ends (ideal for filter coefficients)// Multi-window comparison: Hann vs other windows
const n = 512;
const hann = torch.hann_window(n);
const hamming = torch.hamming_window(n); // Slightly different sidelobe behavior
const blackman = torch.blackman_window(n); // Much lower sidelobes, wider main lobe
// Different windows trade off frequency resolution vs spectral leakageSee Also
- PyTorch torch.hann_window(window_length, periodic=True, dtype=None, layout=torch.strided, device=None, requires_grad=False)
- hamming_window - Similar but slightly different sidelobe behavior
- blackman_window - Lower sidelobes, wider main lobe
- bartlett_window - Triangular, simpler but less optimal
- kaiser_window - Tunable trade-off between main lobe and sidelobes