High Performance Speaker Verification Using Wideband, rich Database

سال انتشار: 1395
نوع سند: مقاله کنفرانسی
زبان: انگلیسی
مشاهده: 578

فایل این مقاله در 5 صفحه با فرمت PDF قابل دریافت می باشد

استخراج به نرم افزارهای پژوهشی:

لینک ثابت به این مقاله:

شناسه ملی سند علمی:

CBCONF01_0234

تاریخ نمایه سازی: 16 شهریور 1395

چکیده مقاله:

Speaker verification has been studied for years. Many databases such as NIST has been used widely ;however , most of these databases are narrow band, not rich in context information and have high channel effect. In this paper, a wide band low noise and rich database of Farsi language has been used which does not have mentioned problems and it is suitable for many applications. Feature extraction is a key part in speaker verification. STFT-MFCC which uses FFT and filter bank is state of the art feature in speaker verification. The main problem of STFT-MFCC is that cannot model envelope accurately. We use STRAIGHT-MFCC, which is well-known for synthesis. STFT-MFCC and STRAIGHT-MFCC performance was compared for 2 minutes and full training data using GMM-UBM model. Results show that STRAIGHT-MFCC outperforms STFT-MFCC especially for short duration training data

کلیدواژه ها:

نویسندگان

Ali Kafaei

Master Student Amirkabir University of Technology Tehran, Iran

Abolghasem Sayadian

Associated Professor Amirkabir University of Technology Tehran, Iran

Hamidreza Baradaran Kashani

PHD Candidate Amirkabir University of Technology Tehran, Iran