Apply filter function

[mthrok]

With file-like object support in StreamReader/Writer, we can add feature to apply arbitral FFmpeg effect to Tensor.

https://docs.google.com/document/d/1qrCWKqc8mjs1lKCvihloJihJhf3KRYCDD0JPCNeJV1s/edit#

Starting code

import torch
from io import BytesIO
import torchaudio
from torchaudio.io import StreamReader, StreamWriter

torchaudio.utils.ffmpeg_utils.set_log_level(56)

class StreamingEncoder:
    def __init__(self, src: torch.Tensor, sample_rate):
        self.src = src

        self.buffer = BytesIO()
        self.writer = StreamWriter(self.buffer, format="f32le")
        self.writer.add_audio_stream(num_channels=src.size(1), sample_rate=sample_rate)
        self.writer.open()

        self.i_read = 0
        self.i_iter = 0
        self.step = 256  # samples

    def read(self, n):
        # refill
        while self.i_iter < self.src.size(0) and self.buffer.tell() - self.i_read < n:
            chunk = self.src[self.i_iter:self.i_iter+self.step]
            self.writer.write_audio_chunk(0, chunk)
            self.i_iter += self.step

        i_write = self.buffer.tell()
        self.buffer.seek(self.i_read)
        ret = self.buffer.read(n)
        self.i_read += len(ret)
        self.buffer.seek(i_write)
        return ret


waveform, sample_rate = torchaudio.load("test/torchaudio_unittest/assets/sinewave.wav")
waveform = waveform.T.contiguous()

fileobj = StreamingEncoder(waveform, sample_rate)

s = StreamReader(fileobj, format="f32le")
s.add_audio_stream(-1, filter_desc="afftfilt=real='hypot(re,im)*sin(0)':imag='hypot(re,im)*cos(0)':win_size=512:overlap=0.75")
s.process_all_packets()
applied, = s.pop_chunks()

import matplotlib.pyplot as plt

f, axes = plt.subplots(2, 1)
axes[0].specgram(waveform[:, 0], Fs=sample_rate)
axes[1].specgram(applied[:, 0], Fs=sample_rate)
plt.show()

[mthrok]

Stab #3163