On the 10th day of Christmas, you'll never guess what my RA gave to me! (🎵Ten Hours of Data🎵)

-

Introduction

In my previous post I showed that SampleRNN seems to perform well with 10-hour datasets. So naturally I wanted to try some more datasets of this length.

Experiments

Birds

The rain dataset from last time produced some nice birdsong. So I found the video shown in Example 1, which contains 10 hours of birdsong.


Example 1: Video containing 10 hours of non-repetitious birdsong, used to train SampleRNN.

I used that to train SampleRNN for 120 000 iterations (a day and a half). Some of the results are in Example 2;



Example 2: Sounds generated by WaveRNN trained for 120 000 iterations on a 10-hour bird dataset.

Fire

I found several long recordings of fire sounds on Youtube, shown in Example 3.




Example 3: Videos containing a total of 12 hours of fire recordings, used to train SampleRNN.

One must be careful when selecting videos from Youtube, because many nominally 10 hour recordings are actually 1 hour of unique audio looped 10 times (BBC has a number of these). Many of these videos are sparse on information about how the recordings were made or amassed. These in places sound like they might have synthesized noise in the background. The first video above may or may not contain a subset of the second video. The third video may or may not contain repetitions. It is actually really difficult to tell, even looking at the waveforms. As a side-note, I was going to do an autocorrelation on the audio in the third video, because a spike would come out telling you whether there is a repetition. However, that would involve a very large DFT, and I'm not sure that is tractable. So, in the interest of time, I just went for it. The results are in Example 4.



Example 4: Sounds generated by WaveRNN trained for 120 000 iterations on a 10-hour fire dataset.

Trains

Will brilliantly suggested 'slow tv' as a source of long audio files. I took the famous Bergen-Oslo train video, which contains about 7 hours of audio recorded on the inside of the train. I mixed that with 45 minutes of the 'train' class in the DCASE 2016 dataset (I truncated each 10-second clip to 8 seconds for training), also from the inside of the train. I additionally included 1 hour from a video recorded on what sounds like the outside of the train. The videos are shown in Example 5.



Example 5: Videos containing a total of 8 hours of train recordings, used to train SampleRNN in conjunction with an additional 45 minutes from DCASE train sounds.

This dataset, in addition to the train sounds, is replete with many very prominent bell sounds and spoken announcements. I trained SampleRNN, and the results are in Example 6.



Example 6: Sounds generated by WaveRNN trained for 120 000 iterations on an 8-hour train dataset.

While these do sound train-like, to me they are leaning back towards sounding like a noise-profile / first-order statistics. They don't have any of the foreground sounds like talking present in the dataset.

Final Apotheosis and Immolation

I think that is the end of training models for this project. I am going to put these samples in the listening tests and launch that online. I might let the listening tests sit and collect data for a few months, and write it up later, maybe in April. I will also finish putting these models in the Audio-By-The-Meter tool, along with some of the other semi-decent models, and transfer that to RPPtv for inclusion in their web-universe.

Comments

Post a Comment

Popular posts from this blog

WaveRNN

-
Image
Figure 1: CVSSP's Pet Wyvern. Introduction I've been thinking about how we can make audio synthesis faster. In particular, I would be interested in realtime soundscape synthesis, partially because I think it would be good for the project, and partly because it would be well aligned with my own personal research goals. I found two options: Parallel Wavenet For a while, I was looking at parallel wavenet . This can supposedly generate samples in realtime, and is now used in the Google Assistant. However, I have been unable to find a good, vanilla implementation of it, and the paper is really sparse on details. There are some projects on GitHub where people have started implementing it more than 6 months ago and have not finished, so given the time constraints of this project, implementing it myself doesn't seem feasible. Moreover, the training process is very intricate and involves training a couple of networks separately and then together -- which makes it really har
Read more

How I calibrated my contact microphone

-
Image
Introduction I used the Hsu-Nielson method to find the frequency response of my contact microphones. This method involves cracking a mechanical pencil graphite on the microphone transducer. The microphone in question is the one shown above, and described in detail here: http://michaelkrzyzaniak.com/marshmallow.php I specifically used the white, low-gain variant. Experimental Setup and Results Figure 1: The mechanical pencil used in this experiment, with 0.5mm number-2 graphite Figure 2: Three millimeters of graphite were extended from the pencil Figure 3: Audio interface used for experiment (Scarlett 2i4), showing the settings used. Figure 4: Video showing the graphite being cracked. A small amount of putty was used to prevent the tip of the pencil tapping the transducer after the graphite breaks. The graphite is placed on the transducer and constant pressure is applied. When the graphite breaks, the pressure is released in a very short time (shorter than th
Read more

Ambisonic Rendering in the Story Bubble

-
Image
Figure 1: Sound Sphere in the Center for Vision, Speech and Signal Processing (CVSSP) at the University of Surrey. The person is shown standing in the intended orientation, facing the front of the sphere. Introduction I recently needed some audio content that I could quickly render for playback in the sound sphere shewn in Figure 1. There are a reasonable amount of free ambisonic recordings available online, e.g. Angelo Farina Youtube So I spent some time putting together tools to render ambisonics in the sphere, and here I wanted to share what I learned in case anybody else at CVSSP wants to try this out in the future. Speaker Arrangement The sphere has 24 speakers (plus a subwoofer which I will henceforth ignore), which at the time of writing are arranged as shown in Figure 2 and Figure 3. Figure 2: Speaker layout. The center of the diagram represents the top of the sphere. The top of the diagram represents the front of the sphere. There is one spe
Read more