# Audio¶

Sound, or audio signals are signals that vibrate in the audible frequency range. When someone talks, it generates air pressure signals; when you play music, the speaker converts electrical energy into mechanical energy. The mechanical energy compresses air and converts the motion into sound energy. Your ear takes in these air pressure differences and communicates with the brain. That’s how audio works in the real world and we call it analog audio.

Digital audio data can be a soundbite, music, a ringtone, a background noise. It often comes in .wav, .mp3 formats, where the sound waves are digitized by sampling them at discrete intervals. There are many interesting neural search tasks on audio data: music recommendations, similarity search for audio files, matching voice commands, voice synthesis for chatbot. Before you start to build solutions with Jina, let’s recap some fundamental knowledge about audio data.

## Sampling rate¶

Sampling rate determines the sound frequency range (corresponding to pitch) which can be represented in the digital waveform. It represents the quality of your audio data. Typical sampling frequencies are 8KHz, 16KHz and 44.1KHz. 1Hz means one sample per second, so obviously higher sampling frequencies mean more samples per second and therefore better signal quality.

Different data loaders, mediums and file formats might have different sample rate requirements. For example, typical studio recording audio has 192KHz. To make this recording as a CD, it should be resampled to CD sampling rate of 44.1KHz.

Fun fact, human ear can hear up to ~20KHz. Higher frequencies convey some emotions, and are useful for identification of the speaker. In particular, if it is a human voice, then it is still intelligible at much lower sampling rate even when higher frequencies are lost.

And you really don’t want to hear a 4KHz music.

## Quantization¶

Another important concept in audio data is quantization. It is the process of reducing the infinite number precision of an audio sample to a finite precision as defined by a particular number of bits. In the majority of the cases, 16 bits per sample are used to represent each quantized sample, which means that there are 216 levels for the quantized signal. Quantization also forms the core of essentially all lossy compression algorithms.

Following image depicts a simple wave quantized in 2bits (left) and 3bits (right).

Let’s listen to some soundbites and feel how quantization affects the audio quality. In this example, the original audio file is at 16-bit, 44.1KHz.

Now, we keep the sampling rate at 44.1KHz but reduce the number of bits used in quantization:

Quantization Audio
8-bit
6-bit
4-bit
3-bit
2-bit

Don’t be flattered if you still recognize the 2-bit version. It is easy only because it is a human voice, which we are very good at recognizing. Try again with a music.

## Load .wav file¶

You can use Jina Document API to load a wav file as a Document.

from jina import Document

print(d.blob.shape, d.blob.dtype)

(30833,) float32


## Save as .wav file¶

You can save Document .blob as a .wav file:

d.dump_audio_blob_to_file('toy.wav')


## Example¶

Let’s load the “hello” audio file, reverse it and finally save it.

from jina import Document

d.blob = d.blob[::-1]
d.dump_audio_blob_to_file('olleh.wav')

hello.wav olleh.wav

## Other tools & libraries for audio data¶

By no means you are restricted to use Jina native methods for audio processing. Here are some command-line tools, programs and libraries to use for more advanced handling of audio data: