Effects of Audio and ASR Quality on Cepstral and High-level Speaker Verification Systems

Speech data for NIST speaker recognition evaluations has traditionally been distributed in compressed, telephone quality form, even for microphone data that was origi-nally recorded at higher quality. We evaluate the effect that improved audio quality has for speaker verification performance, using a recently released full-bandwidth version of microphone data from the SRE2010 evalua-tion. Remarkably, we find substantially improved results even though the underlying speaker recognition models remain based on a telephone-band feature front end. For a cepstral GMM system we show improvements purely from the elimination of lossy (µlaw) coding and more ef-fective noise reduction filtering at the full bandwidth. We also find that higher-level speaker recognition systems can benefit from better ASR quality enabled by the im-proved audio quality. Specifically, we show that a speech recognizer trained on full-bandwidth, distant-microphone meeting speech data yields reduced speaker verification error for speaker models based on MLLR features and word-N-gram features. 1