A little bit of background

Over recent years, i-vectors have become very popular in the field of speaker recognition. i-vectors have also demonstrated superior performance for speaker verification, language recognition and speaker diarization tasks. Each utterance is projected onto a total factor space and is represented by a low-dimensional feature vector.

I-vector Extraction

1. Download files - ivector-create, ivector-compare, BigFile.bin.

2. Run ivector-create.

./ivector-create *.wav (*.wav - speaker's recording). The program creates i-vectors file *.iv.

3. Do the same, but with another WAV file.

4. Run ivector-compare.

./ivector-compare 1.iv 2.iv The output of the program contains cosine distance scoring.

For testing purposes, you can download MP3 files (inaugural addresses of the presidents of the United States) from Microsoft Speaker Recognition API, convert it to WAV, make i-vectors and compare them. (It is recommended to remove the audience applause from WAV).

The utilities were tested on Lubuntu 5.10, 6.04.

If you have questions, please get in touch with us at speech_ua at yahoo.com.