TL;DR
This paper tackles the challenge of creating natural-sounding Amharic speech synthesis by focusing on the crucial role of stress, which is phonetically linked to consonant gemination. Researchers developed a method to generate stressed syllables by manipulating consonant duration and identified specific duration thresholds for voiced and unvoiced consonants, leading to a significant improvement in synthesized speech naturalness.
What Problem Does This Paper Address?
Developing natural and intelligible speech synthesis systems for languages other than English and Japanese remains a significant challenge. Amharic, the official language of Ethiopia, presents a unique hurdle: stress, a key element of prosody and intelligibility, is closely tied to consonant gemination (doubling). While gemination is phonemically meaningful in Amharic, its acoustic correlates and how to effectively generate them in a speech synthesizer were not well-understood. Previous Amharic speech synthesis efforts were limited, and this paper aims to bridge this gap by investigating the acoustic characteristics of Amharic stress and developing a rule-based system to incorporate it.
Methods
The researchers employed a cepstral method for Amharic speech synthesis, building upon a general parametric and rule-based system. The core of their contribution lies in developing a stress generation rule. This involved:
- Acoustic Analysis: Analyzing minimal pair words that differ only by the presence or absence of geminated consonants. They examined time-varying patterns of pitch, power, and spectrum to identify acoustic correlates of stress.
- Synthesis Experiments: Synthesizing speech by manipulating these acoustic parameters (pitch, power, spectrum, and duration) of stressed and unstressed syllables. This allowed them to isolate the impact of each feature.
- Perceptual Listening Tests: Conducting listening tests with native Amharic speakers to determine which acoustic features were most perceptually salient for stress.
- Duration Threshold Determination: Specifically investigating the role of consonant duration by systematically varying it and observing listener perception.
- Stress Generation Rule: Based on their findings, they developed a rule that primarily lengthens the duration of consonants to generate stressed syllables. For continuant consonants, they repeated the first three frames of the consonant part for 90ms. For non-continuant consonants, a 100ms silence was added.
Datasets
The paper does not explicitly mention the creation of new datasets. Instead, it utilizes:
- Minimal Pair Words: A set of 15 minimal pair words were extracted from an Amharic-English dictionary, with 6 pairs used for detailed analysis. These words differ solely in the presence or absence of geminated consonants.
- Synthesized Speech Data: Sixteen types of synthesized speech data were generated for listening tests, including analysis-synthesis sounds and combinations of exchanged acoustic parameters from stressed and unstressed words.
- Original Speech Data: Original utterances of the minimal pair words were also used for comparison.
Key Insights and Ideas
- Consonant Duration is King for Amharic Stress: Contrary to the common assumption that stress is primarily a vowel feature, this study strongly suggests that consonant duration is the most critical acoustic correlate for Amharic stress. Manipulating consonant duration had a much greater perceptual impact than altering pitch, power, or spectrum.
- Specific Duration Thresholds: The research identified distinct duration thresholds for consonants: approximately 50ms for voiced consonants and 70ms for unvoiced consonants, beyond which a syllable is perceived as stressed.
- Gemination as the Physical Manifestation of Stress: The paper posits that gemination in Amharic is the physical phenomenon related to production (duration), while stress is the psychological phenomenon related to perception (prominence).
- Simplicity Yields Naturalness: A straightforward stress generation rule, primarily focused on consonant lengthening, proved highly effective in improving the naturalness of synthesized Amharic speech.
Key Results
- High Intelligibility: The authors report that their previous system achieved 100% correct-rate in intelligibility tests.
- Improved Naturalness: The newly developed stress generation rule significantly improved the naturalness of synthesized stressed words.
- Duration Thresholds:
- Voiced consonants: Average threshold duration of 50ms.
- Unvoiced consonants: Average threshold duration of 70ms.
- Effectiveness of Stress Rule: The stress generation rule achieved a 100% correct-rate in generating stressed words, with exceptions for certain exceptional 6th-order syllables.
Limitations and Future Work
- Limited Stress Identification: The current system requires users to explicitly input stress marks. It lacks an automatic mechanism to identify stressed syllables from input text.
- 6th Order Syllable Exception: The stress rule was not effective for all 6th-order Amharic syllables, which exhibit unique properties (e.g., lack of vowel in unstressed form). This suggests that vowels might be necessary for syllables to be perceived as stressed.
- Future Work:
- Develop a lookup dictionary to automatically identify stressed syllables.
- Investigate the effect of stress on neighboring syllables, particularly vowels preceding geminates.
- Extend the stress generation capabilities to the sentence level.
Summary updated