2010, a speech oddity: phonetic transcription of reversed speech

François Pellegrino, Emmanuel Ferragne, Fanny MeunierInterspeech 2010

TL;DR

This paper investigates how well humans can phonetically transcribe speech played backward, a common experimental control in speech perception research. Surprisingly, expert phoneticians could accurately transcribe many phonemes and even entire pseudo-words, suggesting reversed speech retains significant phonetic information and might not be a purely "non-speech" control.

What Problem Does This Paper Address?

In speech perception research, time-reversed speech (RS) is often used as a control stimulus. However, there's ambiguity about whether RS should be treated as "non-speech" (activating only low-level auditory processing) or as a form of "delexicalized speech" (triggering speech perception mechanisms). This stems from a lack of detailed understanding of how time reversal affects the phonetic cues within speech sounds. Previous studies have primarily focused on word-level intelligibility or used orthographic transcriptions, leaving the phonetic perception of RS largely unexplored. This paper aims to fill this gap by assessing the extent to which phonetic cues are preserved in reversed French pseudo-words and how experts perceive them.

Methods

The core of the study involved expert phonetic transcription of time-reversed French pseudo-words.

  • Stimuli Generation: 47 French pseudo-words were recorded and then globally time-reversed.
  • Transcription Task: Four expert phoneticians, unaware of the study's specific goals, were tasked with providing detailed phonetic transcriptions of these reversed stimuli. They could listen to each stimulus multiple times.
  • Transcription Systems: Three phoneticians used the International Phonetic Alphabet (IPA), while one used SAMPA.
  • Analysis: The transcriptions were analyzed for accuracy in broad phonetic categories (e.g., vowels, stops, fricatives) and for the precise retrieval of original phonetic features and segments. Agreement between subjects was also assessed.

Datasets

The study utilized a custom dataset of 47 French pseudo-words. These were digitally recorded and then time-reversed to create the experimental stimuli. No pre-existing datasets were used for training or evaluation.

Key Insights and Ideas

  • RS is More "Speech-like" Than Expected: The most striking finding is that expert phoneticians could accurately transcribe a significant portion of reversed speech, including entire pseudo-words. This challenges the notion of RS as purely "non-speech."
  • Phonetic Cues Persist: Many phonetic features, particularly for continuant sounds (like fricatives and nasals), are remarkably preserved under time reversal. This suggests that the acoustic information crucial for identifying these sounds is robust to reversal.
  • Asymmetry in Perception: The study hints at perceptual asymmetries, particularly concerning the perception of duration. The authors note that subjects transcribed vowel length and gemination, which aligns with existing research on how ramping vs. damping amplitude envelopes are perceived differently.
  • Stops are Problematic: Unvoiced stops proved to be the most challenging for transcription, often being decomposed into multiple segments or misidentified as fricatives. This highlights the unique spectro-temporal dynamics of stops that are significantly altered by reversal.
  • Potential for Lexical Interference: The high intelligibility of phonetic cues in RS suggests that it could potentially trigger lexical access if the reversed sequence resembles a real word, which has implications for its use as a control in experiments.

Key Results

  • Overall Accuracy: Experts achieved a global detection accuracy of 66.9% for broad phonetic classes across 616 transcribed segments.
  • Perfect Retrieval: For over 25% of the stimuli, experts perfectly retrieved the original segments and the CVC structure of the pseudo-words.
  • High Accuracy for Continuants: Oral vowels (88.6% accuracy) and fricatives (91.7-93.3% accuracy) were identified with high accuracy. Nasals also showed high accuracy (90.0%).
  • Low Accuracy for Stops: Unvoiced stops had very low detection accuracy (9.4%), with transcriptions often being fricatives (29%) or decomposed into clusters. Voiced stops were better (61.8%).
  • Schwa and Nasal Vowel Challenges: Phonetic schwas were often detected but sometimes transcribed as complex sequences. Reversed nasal vowels were frequently transcribed as nasal-vowel sequences.
  • Segment Count: While many segments were accurately transcribed, the number of detected segments sometimes varied from the original, particularly for stimuli containing stops, leading to overestimations in some cases.

Limitations and Future Work

  • Expert Bias: The study relied on expert phoneticians, whose enhanced perceptual abilities might not generalize to naive listeners.
  • Pseudo-words: The use of pseudo-words, while controlling for lexical effects, means the findings might differ for real words.
  • Global Reversal: The study focused on global time reversal. Local reversal techniques, used in some intelligibility studies, might yield different results.
  • Future Directions: The authors suggest further investigation into the neurocognitive mechanisms underlying RS perception, particularly how abnormal transitions might elicit responses like Mismatch Negativity. They also propose exploring the implications of RS's phonetic intelligibility for its use as a control stimulus in various experimental paradigms.

Summary updated

BibTeX
@inproceedings{pellegrino2010speech,
  author = {François Pellegrino and Emmanuel Ferragne and Fanny Meunier},
  title = {{2010, a speech oddity: phonetic transcription of reversed speech}},
  booktitle = {Interspeech},
  year = {2010},
}
Figure