TL;DR
This paper analyzes pronunciation variations in American English using the TIMIT corpus, focusing on phonemic substitutions, deletions, and insertions for both vowels and consonants. The findings reveal distinct patterns: vowels are more prone to substitution (often related to raising and reduction), while consonants are more likely to be deleted. This work contributes to understanding "comprehensible" pronunciation for language learning and improving speech recognition systems.
What Problem Does This Paper Address?
Learning to speak a second language proficiently often involves achieving an "intelligible" pronunciation, rather than a completely accent-less one. This requires not only teaching canonical pronunciations but also accounting for the natural variations that occur in spontaneous speech. While previous research has looked at pronunciation variations at the word level or for speech recognition modeling, this paper aims to provide a more detailed phonemic-level analysis of these variations in American English. Understanding these patterns is crucial for developing effective pronunciation training tools and for building more robust speech recognition systems that can handle real-world speech.
Methods
The researchers analyzed the TIMIT corpus, a widely used dataset of read American English speech. They compared the manually transcribed phonetic pronunciations in TIMIT with the canonical pronunciations defined in the CMU Pronunciation Dictionary. To facilitate this comparison at the phonemic level, the TIMIT phonetic transcriptions were converted to the CMU 39 phoneme set using a set of collapsing rules. The analysis then focused on identifying and quantifying three types of variations: substitution (one phoneme replaced by another), deletion (a phoneme is omitted), and insertion (an extra phoneme is added). These variations were analyzed separately for vowels and consonants.
Datasets
The primary dataset used in this study is the TIMIT corpus of American English. This corpus contains 6,300 read sentences from 630 speakers, totaling approximately 5,100 unique word types and 54,391 word tokens. The canonical pronunciations were sourced from the CMU Pronunciation Dictionary.
Key Insights and Ideas
- Vowels vs. Consonants: Different Strategies: Vowels and consonants exhibit fundamentally different strategies for variation. Vowels are more likely to be substituted, often undergoing changes related to their height and frontness/backness (raising and reduction). Consonants, on the other hand, are more frequently deleted.
- The "Schwa" Phenomenon: Many vowel variations converge towards the schwa sound (/ah/ or /ih/ in the CMU set), suggesting that these reduced vowel sounds are common realizations in spontaneous speech.
- Consonant Variations Tied to Articulation Features: Consonantal substitutions are strongly linked to changes in voicing, place of articulation, and manner of articulation. For instance, voicing changes are common (e.g., /t/ becoming /d/), as are shifts in where the sound is produced in the mouth.
- Alveolar Stops are "Unstable": The alveolar stops /t/ and /d/ appear to be particularly prone to variation, both in terms of substitution and deletion.
- Implications for Intelligibility: Understanding these common variations is key to defining what constitutes "comprehensible" pronunciation, moving beyond the idea of a perfect, accent-less rendition.
Key Results
- Vowel Substitution Dominates: Vowels are more subject to substitution (average rate of 13.84%) than deletion (average rate of 5.70%).
- Consonant Deletion is More Common: Consonants are more likely to be deleted (average rate of 7.62%) than substituted (average rate of 13.84%).
- Specific Vowel Patterns: Vowels like /ah/ and /uw/ show significant substitution rates, with /ah/ being the most prone to reduction.
- Specific Consonant Patterns: Voicing changes are prominent, with /t/ being realized as /d/ by 19.6%. Alveolar stops (/t/, /d/) are identified as the most vulnerable consonants to substitution.
- High Canonical Realization for Consonants: Consonants, on average, maintain their canonical form at a higher rate (86.91%) compared to vowels (79.52%).
Limitations and Future Work
The authors acknowledge that examining the causes of these variations is beyond the scope of this paper and requires future research. They also suggest that future work should focus on determining which of these observed variations fall under the umbrella of "comprehensible" pronunciation. This distinction is crucial for developing practical applications in English pronunciation learning and assessment systems. The study also notes alternations between vowel and consonant sounds that don't fit neatly into the presented tables, indicating areas for further investigation.
Summary updated