Wideband parametric speech synthesis using warped linear prediction

Tuomo Raitio, Antti Suni, Martti Vainio, Paavo AlkuInterspeech 2012

TL;DR

This paper introduces Warped Linear Prediction (WLP) as a more effective method for modeling wideband speech spectra in parametric speech synthesis. By adapting the frequency resolution to better match human hearing, WLP allows for higher quality synthetic speech at increased sampling rates compared to traditional Linear Prediction (LP).

What Problem Does This Paper Address?

Traditional parametric speech synthesis systems typically operate at a 16 kHz sampling rate, providing an 8 kHz audio bandwidth. While this is sufficient for intelligible speech, higher sampling rates (e.g., 44.1 kHz) offer the potential for more natural and higher-fidelity synthetic speech. However, simply increasing the sampling rate with conventional Linear Prediction (LP) presents challenges. LP uses a linear frequency resolution, which becomes inefficient and computationally expensive at higher sampling rates because the number of actual speech resonances doesn't increase proportionally. This leads to a need for higher model orders, which can degrade the accuracy of spectral estimation and statistical modeling. The paper addresses this by exploring Warped Linear Prediction (WLP), which offers a perceptually motivated frequency resolution, aiming to improve spectral modeling accuracy and potentially reduce model order requirements for wideband speech synthesis.

Methods

The core of this paper is the application of Warped Linear Prediction (WLP) to parametric speech synthesis. WLP modifies the standard LP by replacing unit delay elements with first-order all-pass filters. This warping of the frequency scale allows for a non-uniform frequency resolution, giving better resolution at lower frequencies where human hearing is more sensitive and speech formants are most critical.

The authors integrated WLP into their existing HMM-based parametric speech synthesis system, GlottHMM. This system decomposes speech into a vocal tract filter and a voice source. WLP is used to model the vocal tract spectrum, while the voice source is parameterized by energy, fundamental frequency (F0), harmonic-to-noise ratio (HNR), and its own LP spectrum. For training, LP-based parameters are converted to line spectral frequencies (LSF).

The key novelty lies in replacing the conventional LP filters (both FIR and IIR) with their warped counterparts within the synthesis pipeline. The paper also investigates the selection of the warping coefficient ($\lambda$) and the model order, which are crucial for optimizing WLP performance in a synthesis context.

Datasets

The speech material used for training and evaluation was a subset of the Finnish "Heini" database. This database consists of phonetically rich sentences and continuous non-fiction read by a 27-year-old female speaker. The dataset was automatically annotated with word prominence labels and segmented using HTS. For training, full-context labels with quinphones, word prominence, and positional/quantitative features were utilized. For listening tests, randomly chosen sentences from a held-out portion of the database were used.

Key Insights and Ideas

  • WLP offers perceptually better spectral modeling: By warping the frequency scale, WLP aligns spectral resolution more closely with human auditory perception, leading to more accurate spectral models, especially at higher sampling rates.
  • WLP can improve wideband synthesis quality: The experiments demonstrate that WLP-based wideband synthetic speech is rated higher in quality than both narrowband speech and wideband speech synthesized with conventional LP.
  • Model order and warping are crucial tuning parameters: While WLP is beneficial, the optimal warping coefficient ($\lambda$) and model order require careful experimental tuning in the context of parametric speech synthesis, as the synthesis process differs from speech coding.
  • Bandwidth and model order impact quality, but WLP is the primary driver: Increasing bandwidth and model order generally improved quality, but the most significant gains were observed when WLP was employed.
  • Speaker similarity improvements are not guaranteed by WLP alone: While previous studies suggested WLP could improve speaker similarity at higher sampling rates, this study did not find statistically significant differences in similarity scores across the tested systems. This might be due to limitations in the voice source modeling.

Key Results

  • Quality Improvement: WLP-based wideband synthetic speech was rated significantly higher in quality than conventional LP-based wideband speech and 16 kHz narrowband speech.
  • WLP vs. LP: For 44.1 kHz speech with a 30th order model, WLP systems (with $\lambda=0.35$ and $\lambda=0.76$) were rated higher than the conventional LP system. The $\lambda=0.35$ system showed a slight edge.
  • Bandwidth and Model Order: A 44.1 kHz system with a 50th order WLP model ($\lambda=0.35$) yielded the best overall quality score, outperforming a 16 kHz baseline and a 44.1 kHz 50th order LP system.
  • Speaker Similarity: No statistically significant differences in speaker similarity were found between the systems, although the best WLP system showed a trend towards higher similarity.

Limitations and Future Work

The authors acknowledge that the quality improvements, while statistically significant, were relatively small. They attribute this partly to the excitation generation method, specifically the exclusion of HNR modification, which may have made the high-band excitation too periodic, particularly for the female voice. They suggest that LP-based wideband speech was not statistically different from LP-based narrowband speech due to the difficulty of re-estimating high-order models with conventional LP, which can place LSFs in irrelevant frequency ranges.

Future work suggested includes:

  • Improving the excitation generation by incorporating HNR modification.
  • Utilizing a wideband pulse library for more sophisticated excitation generation.
  • Further investigation into the optimal selection of model order and warping coefficient for parametric speech synthesis.
  • Exploring the impact of WLP on speaker similarity with improved voice source modeling.

Summary updated

BibTeX
@inproceedings{raitio2012wideband,
  author = {Tuomo Raitio and Antti Suni and Martti Vainio and Paavo Alku},
  title = {{Wideband parametric speech synthesis using warped linear prediction}},
  booktitle = {Interspeech},
  year = {2012},
}

Loading related content...

Figure