i dZddlZddlZddlZddlmZdejfdZ ddejde de d e fd Z d ejfd Zd ejfdZde de de fdZddejfdZdde de de fdZde de fdZy)z1 Signal processing or PyTorch related utilities. N) functionalxctj|dk(tjd|j|jtj ||z S)zh Implementation of sinc, i.e. sin(x) / x __Warning__: the input is not multiplied by `pi`! rg?)devicedtype)torchwheretensorrrsin)rs `/Volumes/fast/ai/experiments/voice-extract-mac/.venv/lib/python3.12/site-packages/julius/core.pysincr sC ;;qAvu||BqxxqwwOQVQZQZ[\Q]`aQa bbr target_lengthmodevaluecXtj|d||jdz f||S)zI Pad the given tensor to the given length, with 0s on the right. r)rr)Fpadshape)r rrrs r pad_tors- 55!]V\\"-==>TQV WWrfreqsc>dtjd|dz zzS)z Converts a Tensor of frequencies in hertz to the mel scale. Uses the simple formula by O'Shaughnessy (1987). Args: freqs (torch.Tensor): frequencies to convert. # )rlog10)rs r hz_to_melrs  %++a%#+o. ..rmelscdd|dz zdz zS)z Converts a Tensor of mel scaled frequencies to Hertz. Uses the simple formula by O'Shaughnessy (1987). Args: mels (torch.Tensor): mel frequencies to convert. r rr)rs r mel_to_hzr#)s "td{#a' ((rn_melsfminfmaxcttjt|j }ttjt|j }tj |||}t |S)z Return frequencies that are evenly spaced in mel scale. Args: n_mels (int): number of frequencies to return. fmin (float): start from this frequency (in Hz). fmax (float): finish at this frequency (in Hz). )rrr floatitemlinspacer#)r$r%r&lowhighrs r mel_frequenciesr-4s` ELLt- . 3 3 5C U\\%+. / 4 4 6D >>#tV ,D T?rc\tj||dzjdzdzS)z$ Return the volume in dBFS. rr!)rrmean)rfloors r volumer2Es) ;;u1{{2. /" 44rfreqsrdurctjt||z|j|z }tjdt j z|z|zS)z Return a pure tone, i.e. cosine. Args: freq (float): frequency (in Hz) sr (float): sample rate (in Hz) dur (float): duration (in seconds) )rr/)rarangeintr(cosmathpi)r3r4r5rtimes r pure_toner=LsK <<BH f 5 ; ; = BD 99Q[4'$. //r kernel_sizestridect|j}|jd}tjt |||z |z dz}|dz |z|z}t j|d||z fj}g}t|jD]"} |j|j| $|jddk(sJd||dgz}|j|||gz|S)aV1D only unfolding similar to the one from PyTorch. However PyTorch unfold is extremely slow. Given an input tensor of size `[*, T]` this will return a tensor `[*, F, K]` with `K` the kernel size, and `F` the number of frames. The i-th frame is a view onto `i * stride: i * stride + kernel_size`. This will automatically pad the input to cover at least once all entries in `input`. Args: input (Tensor): tensor for which to return the frames. kernel_size (int): size of each frame. stride (int): stride between each frame. Shape: - Inputs: `input` is `[*, T]` - Output: `[*, F, kernel_size]` with `F = 1 + ceil((T - kernel_size) / stride)` ..Warning:: unlike PyTorch unfold, this will pad the input so that any position in `input` is covered by at least one frame. rrrzdata should be contiguous)listrpopr:ceilmaxrr contiguousrangedimappendr? as_strided) inputr>r?rlengthn_frames tgt_lengthpaddedstridesrGs r unfoldrPYs.  E YYr]Fyy#fk2[@FJKaOHQ,&(;6J UU51j612 3 > > @FGVZZ\"v}}S)*# ;;r?a )N)__doc__r:typingtprtorch.nnrrTensorr r8strr(rrr#r-r2r=rPr"rr rZs  $cELLcX5<<XX3XTYX /U\\ /)ELL)CuE"5ell5 0E 0u 0 0!Gs!GC!Gr