Methods and systems for cross-correlating and aligning parallel speech utterances to improve quality assurance

A method for determining quality assurance of parallel speech utterances, the method comprising:

• converting a candidate utterance and a reference utterance in obtained audio data into first and second time series sequence representations, respectively;
• performing a cross-correlation of the first and second time series sequence representations;
• aligning the first and second time series sequence representations based on one or more differences in lengths of the first and second time series sequences representations to generate an aligned version of the first and second time series sequence representations;
• after the alignment, adjusting the aligned version of the first and second time series sequence representations based on a phoneme-based alignment;
• generating an alignment difference of path-based distances between the reference and candidate speech utterances based on the adjusted aligned version of the first and second time series sequences representations;
• generating a quality metric based on the cross-correlation of the first and second time series sequence representations and the generated alignment difference of path-based distances between the reference and candidate speech utterances;
• and
• outputting the generated quality metric, wherein the generated quality metric is indicative of a degree of similarity between the candidate and reference utterances.

One or more non-transitory computer-readable media comprising instructions that, when executed by at least one processor, cause the at least one processor to:

• perform a cross-correlation of first and second time series sequence representations generated from a candidate utterance and a reference utterance, respectively, in obtained audio data;
• align the first and second time series sequence representations to generate an aligned version of the first and second time series sequence representations;
• after the alignment, adjust the aligned version of the first and second time series sequence representations based on a phoneme-based alignment;
• generate an alignment difference of path-based distances between the reference and candidate speech utterances based at least in part on the adjusted aligned version of the first and second time series sequences representations;
• generate a quality metric based on the cross-correlation of the first and second time series sequence representations and the generated alignment difference of path-based distances between the reference and candidate speech utterances;
• and
• output the generated quality metric, wherein the generated quality metric is indicative of a degree of similarity between the candidate and reference utterances.