50 years late: repeating miller-nicely 1955

Andrew Lovitt, Jont B. AllenInterspeech 2006

TL;DR

This paper revisits a foundational 1955 experiment on consonant-vowel confusion in noise, replicating it with modern methods and a new dataset. While confirming some original findings, the replication reveals significant differences in how distinct phonetic features like voicing and place are perceived under noisy conditions, suggesting that the original conclusions about feature robustness may not hold universally.

What Problem Does This Paper Address?

In 1955, Miller and Nicely conducted a landmark experiment analyzing how listeners confuse consonant-vowel (CV) sounds when masked by wideband noise. By collecting confusion matrices at various signal-to-noise ratios (SNRs), they identified patterns suggesting that certain phonetic features (like voicing, nasality, and place of articulation) were more robust to noise than others. These findings have been highly influential, shaping our understanding of speech perception in degraded acoustic environments. However, the original experiment had specific methodological constraints, including live speech presentation, a homogeneous subject pool, and potentially different training protocols. This paper aims to re-evaluate these influential conclusions by replicating the core experiment with a modern setup, a larger and more diverse dataset, and a controlled acoustic environment, thereby testing the generalizability and enduring relevance of Miller and Nicely's original findings.

Methods

The core methodology involved presenting CV sounds masked by wideband noise to human listeners and collecting their responses. The authors replicated the spirit of the Miller-Nicely (MN55) procedure but introduced several modernizations:

  • Stimulus Presentation: Instead of live speech, the MN05 experiment used prerecorded CVs from the LDC nonsense speech corpus.
  • Acoustic Environment: Stimuli were presented in a sound booth through headphones with a controlled maximum presentation level (80 dB SPL).
  • Noise and SNR: Wideband noise was used, and two SNRs were tested: quiet and -15 dB SNR. The -15 dB SNR was chosen because MN55 found the most interesting confusions between -12 and -18 dB SNR.
  • Subject Pool: A larger and more heterogeneous group of 26 English-speaking subjects participated, compared to the smaller, homogeneous group in MN55.
  • Response Collection: A graphical user interface (GUI) was used for responses, allowing subjects to repeat stimuli multiple times and offering a "noise only" option if no sound was discernible.
  • Analysis: Confusion matrices were generated and analyzed. A key analytical technique involved developing new orderings for the confusion matrices that minimized a "taxi-cab" distance metric, aiming to group confusions more effectively than the original MN55 ordering. Information theory, specifically mutual information and relative information transmitted for each distinctive feature, was used to quantify feature robustness.

Datasets

  • MN55 Data: The original confusion matrices and analysis from Miller and Nicely's 1955 experiment.
  • MN05 Data: A newly collected dataset of consonant-vowel (CV) sounds masked by wideband noise. CVs were sourced from the LDC nonsense speech corpus. Eighteen talkers and 26 subjects participated.

Key Insights and Ideas

  • Voicing Feature Robustness Re-evaluated: The original MN55 experiment suggested voicing was a highly robust feature. However, the MN05 data shows voicing errors are much more prevalent, particularly in specific confusion groups, and the relative information transmitted for voicing is significantly lower than in MN55.
  • Place Feature Robustness Increased: Conversely, the place of articulation feature appears to be more robust in the MN05 data compared to MN55. This is a direct reversal of the original findings regarding these two features.
  • Feature Robustness is Context-Dependent: The paper highlights that the robustness of distinctive features is not absolute but depends on the specific CVs involved and their confusion patterns. For instance, in MN05, some CV groups exhibit voicing errors while others exhibit place errors at high SNRs.
  • Original Feature Groupings May Not Explain All Confusions: The traditional distinctive feature groupings (voicing, nasality, etc.) used by Miller and Nicely do not fully explain the observed confusion patterns in either the MN55 or MN05 data, especially when new, optimal orderings for the confusion matrices are derived.
  • Methodological Differences Matter: The authors suggest that differences in training protocols, the use of prerecorded vs. live speech, and the homogeneity of the subject pool might contribute to the observed discrepancies between the two datasets.

Key Results

  • Voicing Feature: Relative information transmitted for voicing is significantly lower in MN05 (e.g., ~0.012 at -18 dB SNR) compared to MN55 (e.g., ~0.021 at -18 dB SNR, and much higher at higher SNRs).
  • Place Feature: Relative information transmitted for place is higher in MN05 (e.g., ~0.012 at -18 dB SNR) compared to MN55 (e.g., ~0.001 at -18 dB SNR, and lower at higher SNRs).
  • Nasality, Frication, Duration: These features showed similar relative information transmission across both MN55 and MN05 datasets, suggesting their robustness is less affected by the experimental variations.
  • Confusion Patterns: New optimal orderings for confusion matrices were derived for both datasets, revealing different primary confusion groups. For example, the /ba/ group and the /za/, /sa/, /fa/, /za/ group showed significant errors in MN05 that were not as prominent in MN55.
  • "Noise Only" Responses: In MN05, "noise only" responses were frequent at low SNRs (45% at -18 dB SNR), indicating subjects' inability to discern speech.

Limitations and Future Work

  • Subject Training and Bias: The authors note that the extensive, potentially specialized training in MN55, and the familiarity between talkers and listeners, might have biased the original results. The MN05 experiment, with less intensive training and a "noise only" option, might offer a different perspective on natural speech perception in noise.
  • Prerecorded vs. Live Speech: The use of prerecorded stimuli in MN05, as opposed to live speech in MN55, is a significant methodological difference that could influence results.
  • Limited SNRs in MN05: The MN05 experiment tested only two SNRs (quiet and -15 dB SNR), whereas MN55 tested a wider range (-18 to +12 dB SNR). This limits direct comparison across the full spectrum of noise levels.
  • Future Work: The authors suggest that further investigation into the impact of training protocols, talker-listener familiarity, and the specific acoustic characteristics of the stimuli could clarify the observed differences. Exploring the role of other distinctive features and their interactions under varying noise conditions remains an avenue for future research.

Summary updated

BibTeX
@inproceedings{lovitt2006years,
  author = {Andrew Lovitt and Jont B. Allen},
  title = {{50 years late: repeating miller-nicely 1955}},
  booktitle = {Interspeech},
  year = {2006},
}
Figure