TL;DR
This paper leverages speech technology tools, specifically Automatic Speech Recognition (ASR) and forced alignment, to investigate phonetic and phonological typology. By analyzing European Portuguese (EP) speech data against Italian and German acoustic models, the study reveals that EP fricatives exhibit characteristics more akin to Germanic languages, suggesting a shift away from typical Romance language voicing systems.
What Problem Does This Paper Address?
Linguistic typology aims to categorize languages based on their structural features. One such feature is the implementation of the laryngeal contrast in obstruent consonants (voiced vs. voiceless). Traditionally, Romance languages are characterized as "true voicing" languages, marking this contrast primarily with the [voice] feature (vocal cord vibration). In contrast, Germanic languages are often described as "aspirating" languages, using the [spread glottis] feature (airflow through the glottis). However, previous small-scale acoustic studies suggested that European Portuguese (EP), a Romance language, might deviate from this pattern, exhibiting a hybrid system where its obstruent voicing profiles resemble those of Germanic languages more than other Romance languages like Italian. This paper addresses the question of whether EP's voicing system is indeed shifting towards a Germanic-like "aspirating" system, and if speech technology methodologies can effectively test these phonetic and phonological theories.
Methods
The researchers employed two primary methodologies derived from speech technology:
- Three-way and Two-way Acoustic Model Choice: They used forced alignment, a technique that aligns audio segments with their corresponding phonetic transcriptions. In this method, the ASR system was presented with combinations of acoustic phone models from European Portuguese (EP), Italian, and German. The system was then tasked with choosing the best-fitting model for each phonemically voiced obstruent in the EP corpus. This allowed them to see which language's acoustic representation was most similar to the EP speech. They conducted a "three-way choice" (EP, Italian, German) and a "two-way choice" (Italian, German, excluding EP models).
- Forced Alignment with Pronunciation Variants: This method involved enriching the EP pronunciation dictionary with variants for obstruent voicing. This means that a phonemically voiced obstruent (e.g., /b/) could be detected by the system as either its voiced realization ([b]) or its voiceless counterpart ([p]). The system then had to choose the best-fitting phone model (voiced or voiceless) for each obstruent. This was performed twice, once using Italian acoustic models and once using German acoustic models. This approach allows for testing the similarity of EP obstruents to either Italian or German by observing which language's voiced/voiceless variants are preferentially selected.
Datasets
The primary dataset used was a European Portuguese (EP) corpus consisting of 114 hours of broadcast news speech. This corpus was force aligned using existing corpora acquired from LDC, ELRA, and international projects. For training the acoustic models, comparable datasets were used: EP (1.1 million word tokens), Italian (1.8 million word tokens), and German (1.8 million word tokens).
Key Insights and Ideas
- EP Fricatives Lean Germanic: The most striking finding is that EP fricatives consistently show patterns more similar to German than to Italian. This strongly suggests that the [spread glottis] feature is the primary mechanism for voicing contrast in EP fricatives, aligning with Germanic languages rather than the typical Romance [voice] feature.
- EP Stops are Ambiguous: Unlike fricatives, the voicing patterns for EP stops are less clear-cut. While some results suggest a similarity to Italian, others are inconclusive or even contradictory depending on the method used and the specific place of articulation.
- Hybrid Voicing System for EP: The differing patterns for stops and fricatives support the hypothesis that EP exhibits a hybrid voicing system, as previously suggested by Ramsammy and Strycharczuk (2016). This means EP might use different phonetic features ([spread glottis] for fricatives, and potentially [voice] or a mix for stops) to mark the voicing contrast, unlike the more uniform systems found in typical Romance or Germanic languages.
- Speech Technology as a Linguistic Tool: The study successfully demonstrates the utility of speech technology methodologies (ASR, forced alignment) for investigating phonetic and phonological theories, offering a scalable and data-driven approach to linguistic analysis.
Key Results
- Three-way Model Choice (Section 3): While the original EP acoustic models were generally preferred, for the remaining cases, German models were preferred for fricatives (labiodental /v/ and coronal /z/), and Italian models were preferred for bilabial and velar stops.
- Two-way Model Choice (Italian vs. German) (Section 4): For fricatives, German models were markedly preferred for both labiodental /v/ and coronal /z/. For stops, Italian models were preferred for coronals and velars, while German models were preferred for bilabials.
- Forced Alignment with Pronunciation Variants (Section 5):
- Fricatives: Italian voiceless variants were selected more often than German ones for both labiodental /v/ (Italian 90% vs. German 72%) and coronal /z/ (Italian 57% vs. German 30%). This result is interpreted as EP fricatives being more similar to German, as the partial devoicing in EP would be interpreted as voicelessness by Italian models.
- Stops: Results were more ambiguous. Italian voiceless variants were selected more for bilabials and velars, but not for coronals. This contradicted predictions and previous findings, suggesting coronals patterned differently.
Limitations and Future Work
The authors acknowledge that their methodologies are not direct replications of previous acoustic studies, which used more granular measures like VOT, intensity, and airflow. The current study relies solely on acoustic features incorporated into ASR models (log energy, pitch, cepstral coefficients). This methodological difference might explain some of the diverging results, particularly for stops.
Future work suggested includes:
- Performing acoustic analyses on a subset of the data to better understand the factors at play.
- Incorporating other acoustic correlates of voicing, such as phonetic/phonological context and prosodic information.
- Investigating the role of stress patterns, as EP is believed to be stress-timed like Germanic languages, which could influence voicing.
Summary updated