Mandarin speech pattern discovery using segmental dynamic time warping and posteriorgram features

Speech pattern discovery aims to identify repeated patterns (e.g., word-like units) from speech. This study analyzes speech patterns in a Mandarin speech corpus using segmental dynamic time warping (SDTW). Mel frequency cepstral coefficients (MFCCs) have not been effective for pattern discovery in multi-speaker conditions. The phoneme posteriorgram features are used here in a template-based method. Tests show that phoneme posteriorgram is significantly better than MFCCs for both single- and multi-speaker conditions. The performance upper-bound of SDTW is also investigated when boundary information is available with the segments divided by word boundaries. The results show that the boundaries significantly improve the pattern discovery in terms of both accuracy and efficiency.