Skip to content

Latest commit

 

History

History
81 lines (56 loc) · 3.34 KB

README.md

File metadata and controls

81 lines (56 loc) · 3.34 KB
Raw

Audio landmark fingerprinting as a Node Stream module

This module is a duplex stream (instance of stream.Transform) that converts a PCM audio signal into a series of audio fingerprints. It works with audio tracks as well as with unlimited audio streams, e.g. broadcast radio.

It is one of the foundations of the Adblock Radio project.

Credits

The acoustic fingerprinting technique used here is the landmark algorithm, as described in the Shazam 2003 paper. The implementation in codegen_landmark.js has been inspired by the MATLAB routine of D. Ellis "Robust Landmark-Based Audio Fingerprinting" (2009). One significant difference with Ellis' implementation is that this module can handle unlimited audio streams, e.g. radio, and not only finished audio tracks.

Note the existence of another good landmark fingerprinter in Python, dejavu.

Description

In a nutshell,

  • a spectrogram is computed from the audio signal
  • significant peaks are chosen in this time-frequency map. a latency of 250ms is used to determine if a peak is not followed by a bigger peak.
  • fingerprints are computed by linking peaks with dt, f1 and f2, ready to be inserted in a database or to be compared with other fingerprints.

Spectrogram, peaks and pairs

In the background, about 12s of musical content is represented as a spectrogram (top frequency is about 10kHz). The blue marks are the chosen spectrogram peaks. Grey lines are peaks pairs that each lead to a fingerprint.

Threshold and peaks

Given the same audio, this figure shows the same peaks and the internal forward threshold that prevent peaks from being too close in time and frequency. The backward threshold selection is not represented here.

Usage

npm install stream-audio-fingerprint

The algorithm is in codegen_landmark.js.

A demo usage is proposed in codegen_demo.js. It requires the executable ffmpeg to run.

var decoder = require('child_process').spawn('ffmpeg', [
	'-i', 'pipe:0',
	'-acodec', 'pcm_s16le',
	'-ar', 22050,
	'-ac', 1,
	'-f', 'wav',
	'-v', 'fatal',
	'pipe:1'
], { stdio: ['pipe', 'pipe', process.stderr] });
process.stdin.pipe(decoder.stdin);

var Codegen = require("stream-audio-fingerprint");
var fingerprinter = new Codegen();
decoder.stdout.pipe(fingerprinter);

fingerprinter.on("data", function(data) {
	for (var i=0; i<data.tcodes.length; i++) {
		console.log("time=" + data.tcodes[i] + " fingerprint=" + data.hcodes[i]);
	}
});

and then we pipe audio data, either a stream or a file

curl http://radiofg.impek.com/fg | nodejs codegen_demo.js
cat awesome_music.mp3 | nodejs codegen_demo.js

on Windows:

type awesome_music.mp3 | node codegen_demo.js

Integration in your project

Matching fingerprints in a database is not a trivial topic, I should write a technical note about it some day.

For a reference implementation you can have a look at the code of the Adblock Radio algorithm to catch ads https://github.com/adblockradio/adblockradio/blob/master/predictor-db/hotlist.js#L150.

License

See LICENSE file.