AI solution that can compare a 10 second selfie video against a database of similar videos and determine uniqueness
Budget: $750 – $1,500 USD
I'm looking for a talented freelancer to build an AI solution that can compare a new 10 second selfie video against a private encrypted database of similar videos and determine some simple results/outputs from video and audio.
Elements:
1. The video selfie in portrait mode - person has to be decently placed within the video frame and decently illuminated.
2. The audio from reading of a random short phrase (must be readable 3-5 seconds) in the native language of the person (language can be selected from user input). Phrase can be randomly generated by AI, must never be the same phrase and the phrase prompted must match what the person reads, so audio must be analyzed.
If the subject head is framed properly and illuminated properly, recording will start automatically.
Within the 10 second recording, the person making the selfie will be prompted to read the random phrase out loud (in native language).
The sound must be analyzed in real time so that the phrase read by the human is converted from speech to text and the output must match the sentence prompted by 90% accuracy or more, or he/she has to start all over again.
RESULT:
The result of each new comparison initiated when a new selfie video taken is compared against this database has to be an answer to these 2 simple cascading questions.
1. Is the subject in the video a human being? true or false - accuracy must be over 90% - cannot be fooled by manikins or by very obvious recordings played on another screen
^ this is required before saving the video to the encrypted database.
2. Is the subject a different human compared to the subjects from all other videos by analyzing both video image for face ID and sound for vocal timbre? if it's not different, must output all matches by @username value.
^ this is also required before saving the video to the encrypted database.
I am open for suggestions to increase accuracy and prevent this system from being fooled/hacked.
Elements:
1. The video selfie in portrait mode - person has to be decently placed within the video frame and decently illuminated.
2. The audio from reading of a random short phrase (must be readable 3-5 seconds) in the native language of the person (language can be selected from user input). Phrase can be randomly generated by AI, must never be the same phrase and the phrase prompted must match what the person reads, so audio must be analyzed.
If the subject head is framed properly and illuminated properly, recording will start automatically.
Within the 10 second recording, the person making the selfie will be prompted to read the random phrase out loud (in native language).
The sound must be analyzed in real time so that the phrase read by the human is converted from speech to text and the output must match the sentence prompted by 90% accuracy or more, or he/she has to start all over again.
RESULT:
The result of each new comparison initiated when a new selfie video taken is compared against this database has to be an answer to these 2 simple cascading questions.
1. Is the subject in the video a human being? true or false - accuracy must be over 90% - cannot be fooled by manikins or by very obvious recordings played on another screen
^ this is required before saving the video to the encrypted database.
2. Is the subject a different human compared to the subjects from all other videos by analyzing both video image for face ID and sound for vocal timbre? if it's not different, must output all matches by @username value.
^ this is also required before saving the video to the encrypted database.
I am open for suggestions to increase accuracy and prevent this system from being fooled/hacked.