Media Stream API - Pt. 2

This tutorial is the continuation of Media Stream API - Pt.1.

This shows how to use StreamReader for

• Device inputs, such as microphone, webcam and screen recording

• Generating synthetic audio / video

• Applying preprocessing with custom filter expressions

import torch
import torchaudio

print(torch.__version__)
print(torchaudio.__version__)

Out:

1.12.0
0.12.0

try:
    from torchaudio.io import StreamReader
except ModuleNotFoundError:
    try:
        import google.colab

        print(
            """
            To enable running this notebook in Google Colab, install nightly
            torch and torchaudio builds and the requisite third party libraries by
            adding the following code block to the top of the notebook before running it:

            !pip3 uninstall -y torch torchvision torchaudio
            !pip3 install --pre torch torchaudio --extra-index-url https://download.pytorch.org/whl/nightly/cpu
            !add-apt-repository -y ppa:savoury1/ffmpeg4
            !apt-get -qq install -y ffmpeg
            """
        )
    except ModuleNotFoundError:
        pass
    raise

import IPython
import matplotlib.pyplot as plt

base_url = "https://download.pytorch.org/torchaudio/tutorial-assets"
AUDIO_URL = f"{base_url}/Lab41-SRI-VOiCES-src-sp0307-ch127535-sg0042.wav"
VIDEO_URL = f"{base_url}/stream-api/NASAs_Most_Scientifically_Complex_Space_Observatory_Requires_Precision-MP4.mp4"

Audio / Video device input

See also

• Accelerated Video Decoding with NVDEC.

• Online ASR with Emformer RNN-T.

• Device ASR with Emformer RNN-T.

Given that the system has proper media devices and libavdevice is configured to use the devices, the streaming API can pull media streams from these devices.

To do this, we pass additional parameters format and option to the constructor. format specifies the device component and option dictionary is specific to the specified component.

The exact arguments to be passed depend on the system configuration. Please refer to https://ffmpeg.org/ffmpeg-devices.html for the detail.

The following example illustrates how one can do this on MacBook Pro.

First, we need to check the available devices.

$ ffmpeg -f avfoundation -list_devices true -i ""
[AVFoundation indev @ 0x143f04e50] AVFoundation video devices:
[AVFoundation indev @ 0x143f04e50] [0] FaceTime HD Camera
[AVFoundation indev @ 0x143f04e50] [1] Capture screen 0
[AVFoundation indev @ 0x143f04e50] AVFoundation audio devices:
[AVFoundation indev @ 0x143f04e50] [0] MacBook Pro Microphone

We use FaceTime HD Camera as video device (index 0) and MacBook Pro Microphone as audio device (index 0).

If we do not pass any option, the device uses its default configuration. The decoder might not support the configuration.

>>> StreamReader(
...     src="0:0",  # The first 0 means `FaceTime HD Camera`, and
...                 # the second 0 indicates `MacBook Pro Microphone`.
...     format="avfoundation",
... )
[avfoundation @ 0x125d4fe00] Selected framerate (29.970030) is not supported by the device.
[avfoundation @ 0x125d4fe00] Supported modes:
[avfoundation @ 0x125d4fe00]   1280x720@[1.000000 30.000000]fps
[avfoundation @ 0x125d4fe00]   640x480@[1.000000 30.000000]fps
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  ...
RuntimeError: Failed to open the input: 0:0

By providing option, we can change the format that the device streams to a format supported by decoder.

>>> streamer = StreamReader(
...     src="0:0",
...     format="avfoundation",
...     option={"framerate": "30", "pixel_format": "bgr0"},
... )
>>> for i in range(streamer.num_src_streams):
...     print(streamer.get_src_stream_info(i))
SourceVideoStream(media_type='video', codec='rawvideo', codec_long_name='raw video', format='bgr0', bit_rate=0, width=640, height=480, frame_rate=30.0)
SourceAudioStream(media_type='audio', codec='pcm_f32le', codec_long_name='PCM 32-bit floating point little-endian', format='flt', bit_rate=3072000, sample_rate=48000.0, num_channels=2)

Synthetic source streams

As a part of device integration, ffmpeg provides a “virtual device” interface. This interface provides synthetic audio / video data generation using libavfilter.

To use this, we set format=lavfi and provide a filter description to src.

The detail of filter description can be found at https://ffmpeg.org/ffmpeg-filters.html

Audio Examples

Sine wave

https://ffmpeg.org/ffmpeg-filters.html#sine

StreamReader(src="sine=sample_rate=8000:frequency=360", format="lavfi")

Signal with arbitral expression

https://ffmpeg.org/ffmpeg-filters.html#aevalsrc

# 5 Hz binaural beats on a 360 Hz carrier
StreamReader(
    src=(
        'aevalsrc='
        'sample_rate=8000:'
        'exprs=0.1*sin(2*PI*(360-5/2)*t)|0.1*sin(2*PI*(360+5/2)*t)'
    ),
    format='lavfi',
 )

Noise

https://ffmpeg.org/ffmpeg-filters.html#anoisesrc

StreamReader(src="anoisesrc=color=pink:sample_rate=8000:amplitude=0.5", format="lavfi")

Video Examples

Cellular automaton

https://ffmpeg.org/ffmpeg-filters.html#cellauto

StreamReader(src=f"cellauto", format="lavfi")

Mandelbrot

https://ffmpeg.org/ffmpeg-filters.html#cellauto

StreamReader(src=f"mandelbrot", format="lavfi")

MPlayer Test patterns

https://ffmpeg.org/ffmpeg-filters.html#mptestsrc

StreamReader(src=f"mptestsrc", format="lavfi")

John Conway’s life game

https://ffmpeg.org/ffmpeg-filters.html#life

StreamReader(src=f"life", format="lavfi")

Sierpinski carpet/triangle fractal

https://ffmpeg.org/ffmpeg-filters.html#sierpinski

StreamReader(src=f"sierpinski", format="lavfi")

Custom filters

When defining an output stream, you can use add_audio_stream() and add_video_stream() methods.

These methods take filter_desc argument, which is a string formatted according to ffmpeg’s filter expression.

The difference between add_basic_(audio|video)_stream and add_(audio|video)_stream is that add_basic_(audio|video)_stream constructs the filter expression and passes it to the same underlying implementation. Everything add_basic_(audio|video)_stream can be achieved with add_(audio|video)_stream.

Note

• When applying custom filters, the client code must convert the audio/video stream to one of the formats that torchaudio can convert to tensor format. This can be achieved, for example, by applying format=pix_fmts=rgb24 to video stream and aformat=sample_fmts=fltp to audio stream.

• Each output stream has separate filter graph. Therefore, it is not possible to use different input/output streams for a filter expression. However, it is possible to split one input stream into multiple of them, and merge them later.

Audio Examples

# fmt: off
descs = [
    # No filtering
    "anull",
    # Apply a highpass filter then a lowpass filter
    "highpass=f=200,lowpass=f=1000",
    # Manipulate spectrogram
    (
        "afftfilt="
        "real='hypot(re,im)*sin(0)':"
        "imag='hypot(re,im)*cos(0)':"
        "win_size=512:"
        "overlap=0.75"
    ),
    # Manipulate spectrogram
    (
        "afftfilt="
        "real='hypot(re,im)*cos((random(0)*2-1)*2*3.14)':"
        "imag='hypot(re,im)*sin((random(1)*2-1)*2*3.14)':"
        "win_size=128:"
        "overlap=0.8"
    ),
]
# fmt: on

sample_rate = 8000

streamer = StreamReader(AUDIO_URL)
for desc in descs:
    streamer.add_audio_stream(
        frames_per_chunk=40000,
        filter_desc=f"aresample={sample_rate},{desc},aformat=sample_fmts=fltp",
    )

chunks = next(streamer.stream())


def _display(i):
    print("filter_desc:", streamer.get_out_stream_info(i).filter_description)
    _, axs = plt.subplots(2, 1)
    waveform = chunks[i][:, 0]
    axs[0].plot(waveform)
    axs[0].grid(True)
    axs[0].set_ylim([-1, 1])
    plt.setp(axs[0].get_xticklabels(), visible=False)
    axs[1].specgram(waveform, Fs=sample_rate)
    return IPython.display.Audio(chunks[i].T, rate=sample_rate)

Original

_display(0)

Out:

filter_desc: aresample=8000,anull,aformat=sample_fmts=fltp

Highpass / lowpass filter

_display(1)

Out:

filter_desc: aresample=8000,highpass=f=200,lowpass=f=1000,aformat=sample_fmts=fltp

FFT filter - Robot 🤖

_display(2)

Out:

filter_desc: aresample=8000,afftfilt=real='hypot(re,im)*sin(0)':imag='hypot(re,im)*cos(0)':win_size=512:overlap=0.75,aformat=sample_fmts=fltp

FFT filter - Whisper

_display(3)

Out:

filter_desc: aresample=8000,afftfilt=real='hypot(re,im)*cos((random(0)*2-1)*2*3.14)':imag='hypot(re,im)*sin((random(1)*2-1)*2*3.14)':win_size=128:overlap=0.8,aformat=sample_fmts=fltp

Video Examples

# fmt: off
descs = [
    # No effect
    "null",
    # Split the input stream and apply horizontal flip to the right half.
    (
        "split [main][tmp];"
        "[tmp] crop=iw/2:ih:0:0, hflip [flip];"
        "[main][flip] overlay=W/2:0"
    ),
    # Edge detection
    "edgedetect=mode=canny",
    # Rotate image by randomly and fill the background with brown
    "rotate=angle=-random(1)*PI:fillcolor=brown",
    # Manipulate pixel values based on the coordinate
    "geq=r='X/W*r(X,Y)':g='(1-X/W)*g(X,Y)':b='(H-Y)/H*b(X,Y)'"
]
# fmt: on

streamer = StreamReader(VIDEO_URL)
for desc in descs:
    streamer.add_video_stream(
        frames_per_chunk=30,
        filter_desc=f"fps=10,{desc},format=pix_fmts=rgb24",
    )

streamer.seek(12)

chunks = next(streamer.stream())


def _display(i):
    print("filter_desc:", streamer.get_out_stream_info(i).filter_description)
    _, axs = plt.subplots(1, 3, figsize=(8, 1.9))
    chunk = chunks[i]
    for j in range(3):
        axs[j].imshow(chunk[10 * j + 1].permute(1, 2, 0))
        axs[j].set_axis_off()
    plt.tight_layout()
    plt.show(block=False)

Original

_display(0)

Out:

filter_desc: fps=10,null,format=pix_fmts=rgb24

Mirror

_display(1)

Out:

filter_desc: fps=10,split [main][tmp];[tmp] crop=iw/2:ih:0:0, hflip [flip];[main][flip] overlay=W/2:0,format=pix_fmts=rgb24

Edge detection

_display(2)

Out:

filter_desc: fps=10,edgedetect=mode=canny,format=pix_fmts=rgb24

Random rotation

_display(3)

Out:

filter_desc: fps=10,rotate=angle=-random(1)*PI:fillcolor=brown,format=pix_fmts=rgb24

Pixel manipulation

_display(4)

Out:

filter_desc: fps=10,geq=r='X/W*r(X,Y)':g='(1-X/W)*g(X,Y)':b='(H-Y)/H*b(X,Y)',format=pix_fmts=rgb24