There are many projects related to ours in the world, for instance,
Siri and Alexa. There are also some projects related to voice lock
and call steering service in most banks, which are more similar to
our project because we are verifying the voice of the customer rather
than talking to them.
The academic research on speaker recognition
(SR) was not carried out until an abduction and murderous case in
1932. Charles Lindbergh, parent of the victim, coincidentally heard
the voice from the criminal near the place where he was asked to
place the ransom, so that he was able to testify against the suspect,
Bruno Hauptmann, over two years later
[1].
This judicial case initiated the first documented research on the
reliability of earwitness by Frances McGehee
[2].
After that, a lot of studies appeared in order to improve the idea
or have different versions of it. Everyone wanted this invention
to work because this idea can open so many opportunities in the
future, which can make so many tasks much faster.
| Feature | Fingerprint | Palmprint | Retina | Iris | Face | Vein | Voiceprint |
|---|---|---|---|---|---|---|---|
| Easy to use | High | High | Low | Medium | Medium | Medium | High |
| Accuracy | High | High | High | High | High | High | High |
| Cost | High | Very High | Very High | Very High | High | Very High | Low |
| User Acceptance | Medium | Medium | Medium | Medium | Medium | Medium | High |
| Remote Authentication | Available | Available | Available | Available | Available | Available | Yes |
| Mobile Phone Collection | Partly Available | Yes | Available | Available | Yes | Available | Yes |
Our team did a literature review according to our clients'
requirements. It involves most of our research outcome.
We have researched and introduced alternive algorithms
and solutions and everything that we found, for instance,
feature extraction methods, and modeling methods, were
included in the document provided below.
We outlined the principles of speaker recognition (SR)
technologies and the differences between speaker
identification and verification at the beginning.
The origin and development of SR approaches as well
as the-state-of-art method for speaker verification
are also involved. Additionally, we described two
options for datasets gathering. Furthermore, we
obtained some feature extraction method and introduced
them briefly in terms of basic ideas. Next, we introduced
feature matching or modeling algorithms. Finally, the
common frameworks for the whole SR system were listed.
For alternive languages, libraries, APIs and frameworks, we didn't compare them in detail. We use python language, some open source libraries and APIs, django framework according to our clients' requirements. These options are all very common in machine learning and webapp development.