Automatic Measurement of Pre-aspiration

Yaniv Sheena, Mísa Hejná, Yossi Adi, Joseph Keshet

This research introduces the first automated tools for measuring pre-aspiration, a significant advancement in laboratory phonology. By employing machine learning, the researchers developed a system capable of matching human-level precision in identifying subtle speech patterns, potentially revolutionizing how linguists analyze language variation.

The Challenge of Pre-aspiration

In phonetics, pre-aspiration refers to a period of "glottal friction"—a breathy sound—that occurs before a voiceless consonant (an obstruent), such as the /h/ sound sometimes heard in the word "kit" or "milk" in specific dialects.

Historically, identifying the exact onset ($t_s$) and offset ($t_e$) of this sound has been a significant hurdle for linguists. The process typically requires subjective, labor-intensive manual annotation of audio waveforms. This time-consuming nature often limits studies to small datasets and introduces risks of human error and inter-transcriber inconsistency. Automating this process allows for larger-scale, objective linguistic research.

Waveform showing the onset ($t_s$) and offset ($t_e$) of a pre-aspiration interval.

Methodology: Structured vs. Frame-based Models

The researchers evaluated two distinct machine learning architectures:

  1. Frame-based Model (Feedforward Neural Network): This model analyzes the speech signal millisecond by millisecond, classifying each individual "frame" as either part of a pre-aspiration event or not. It utilizes a small window of surrounding frames for context and employs post-processing to smooth the results into a single segment.
  2. Structured Prediction Model: Rather than analyzing frames in isolation, this model treats the entire speech segment as a single unit. It uses "feature maps" designed to capture linguistic knowledge—such as the drop in high-frequency energy when a vowel transitions to a consonant. It is trained to minimize "task loss," specifically the timing difference between predicted and actual boundaries.

Both models utilized eight acoustic features, including spectral energy, pitch, voicing detection, and zero-crossing rates.

Experimental Setup and Datasets

The models were trained and tested on two primary datasets:

  • Aberystwyth English: 5,297 manually annotated examples from 16 speakers in Wales.
  • Welsh Language: 607 speech intervals from 16 speakers to test cross-linguistic generalization.

Evaluation was conducted using 5-fold cross-validation and "leave-one-speaker-out" (LOSO) testing to ensure the models were not simply memorizing specific voices.

Results and Linguistic Replication

The structured prediction model significantly outperformed the neural network. Because pre-aspiration boundaries are defined by transitions in the signal (e.g., energy drops), the structured model's ability to analyze the entire segment at once proved superior to the frame-by-frame approach.

Key performance metrics include:

  • Precision: The structured model predicted the onset of pre-aspiration with a mean error of just 2.6 milliseconds.
  • Accuracy: 56.2% of predictions were within 5ms of human expert labels, compared to 43.3% for the frame-based model.
  • Correlation: A very strong positive correlation (r = 0.87) was observed between manual and automated measurements.
Correlation between manually coded and predicted pre-aspiration durations (raw).

Crucially, the automated tool successfully replicated established scientific findings. It correctly identified that pre-aspiration duration is longer following the vowel /a/ and increases as the consonant's place of articulation moves further back in the mouth (e.g., /k/ vs /p/). Furthermore, the model trained on English generalized well to Welsh with only a minor 7.2% drop in accuracy, suggesting the acoustic signature of pre-aspiration is relatively universal.

Future Outlook

While this study represents a major milestone, future work will focus on expanding the model's capabilities. Current limitations include a focus on plosive contexts (stops like p, t, k) rather than fricatives (s, f) and a reliance on high-quality recordings. Future iterations aim to incorporate automatic detection—determining if pre-aspiration occurs before measuring it—and testing performance on lower-quality field recordings.

Summary updated

BibTeX
@inproceedings{sheena2017automatic,
  author = {Yaniv Sheena and Mísa Hejná and Yossi Adi and Joseph Keshet},
  title = {{Automatic Measurement of Pre-aspiration}},
  booktitle = {Interspeech},
  year = {2017},
  doi = {10.21437/interspeech.2017-870},
}
Figure