Data selection for naturalness in HMM-based speech synthesis

Abstract

Can we identify which utterances in a corpus are the best to use for voice training, based on acoustic/prosodic features, and which utterances should be excluded because they will introduce noise, artifacts, or inconsistency into the voice? Can we use found data such as radio broadcast news to build HMM-based synthesized voices? Can we select a subset of training utterances from a corpus of found data to produce a better voice than one trained on all of the data? Which voice training and modeling approaches work best for this type of data? Data and Tools Boston University Radio News Corpus (BURNC): 7+ hours of professionally-read radio broadcast news from 3 female and 4 male speakers Challenges: Multiple speakers, non-TTS speaking style

Adaptation and Frontend Features to Improve Naturalness in Found-Data Synthesis

“…So we posted tiebreaker HITs on MTurk. The tie was not resolved for articulation, so we picked the low setting, corresponding with our prior findings [8,9] that training on hypo-articulated utterances tends to produce better voices. For standard deviation of energy, the low setting was slightly preferred.…”

“…We are specifically interested in radio broadcast news not only because it contains large amounts of speech from each anchor and is often professionally recorded, but because it is available in many languages. In our previous work [8,9], we explored data selection and outlier removal at the utterance level to produce voices that are rated as more natural, even though they were trained on a smaller amount of data than a baseline trained on all of the data. We selected our subsets based on a number of different acoustic and prosodic features, finding that removing outliers for hyper-articulation and combining filters for hypoarticulation and low mean f0 produced voices rated as significantly more natural.…”

Perceptual Ratings of Voice Likability Collected Through In-Lab Listening Tests vs. Mobile-Based Crowdsourcing

“…Other studies using CS for collecting rates on a Likert scale concentrate on audio quality assessments [16,7] employing the discrete 5-point absolute-category-rating (ACR) for subjective Mean Opinion Scores), on voice naturalness [17], and on perceived Quality of Experience (QoE) in a teleconference system [18]. Different to using a Likert scale, the approach in [19] was to ask CS participants to rate correct or incorrect realizations of the /r/ sound in words.…”

“…Paired-comparison approaches have been widely adopted in CS for quality assessment of image [21], video [22], audio [23] and synthetic speech [24,17]. The study in [23] introduces a pair-comparison framework for quantifying QoE of multimedia content as a more convenient approach compared to 5-point scale ratings.…”

“…The study in [23] introduces a pair-comparison framework for quantifying QoE of multimedia content as a more convenient approach compared to 5-point scale ratings. Also the authors of [17] claimed that employing an ACR scale on which to rate speech naturalness "might have been overwhelming" for the raters and therefore conducted later a paired-comparison test, which provided more substantial differences among the tested voices.…”