Identifying an Aurally Distinct Phrase Set for Text Entry Techniques

Document Title

Identifying an Aurally Distinct Phrase Set for Text Entry Techniques

Document Information

• Subject Area: Human-Computer Interaction, Non-Visual Text Entry Systems, Text-to-Speech Technology
• Keywords: text entry, accessibility, text-to-speech (TTS), synthesized speech, input techniques, phrases corpus

Research Background and Problem

• Identified Issues or Challenges:

  • For blind or low-vision users, traditional text entry tasks require visual interaction, limiting accessibility.
  • Existing phrase sets used in text entry research do not ensure clarity when provided via speech.
  • There is a lack of a standardized phrase corpus to test non-visual text entry techniques.

• Significance:

  • Non-visual text entry techniques enable blind or low-vision users to input text more efficiently on portable devices.
  • An accurate phrase corpus is fundamental for evaluating the performance of such techniques and designing solutions to enhance usability.

• Research Motivation and Related Work:

  • Investigating how to create or identify aurally distinct phrase sets for use in the non-visual text entry domain.
  • Building on existing text entry phrase sets (e.g., MacKenzie and Soukoref's 500-phrase set) to optimize their intelligibility in speech scenarios.

Solution

• Proposed Method or Solution:

  • The research team utilized an existing phrase set, generating audio files using text-to-speech systems from Microsoft and Apple.
  • Listening and transcription experiments were conducted with participants recruited from Amazon Mechanical Turk and the target population (blind or low-vision users) to identify aurally distinct phrases.

• Innovations:

  • Proposed a data-driven phrase selection method based on speech generation and human participation.
  • Identified error types in phrases (e.g., spelling errors, spacing errors, comprehension errors) to establish guidelines for future corpus design.
  • Developed a corpus of 92 phrases with low comprehension error rates.

• Implementation Steps and Techniques:

  • Speech Generation: Generated audio files for 500 phrases using Apple macOS and Microsoft Windows text-to-speech technologies, with both male and female voices.
  • Experiment 1:
    • Recruited participants from Amazon Mechanical Turk, collecting approximately 17,421 transcription data points.
    • Analyzed errors using Levenshtein distance and manually categorized error types.
    • Identified 96 phrases with no comprehension errors.
  • Experiment 2:
    • Recruited 80 blind or low-vision users for auditory testing.
    • Validated the auditory clarity of the phrases, resulting in a final set of 92 phrases with no comprehension errors.

Research Outcomes

• Specific Outcomes:

  • Identified a set of 92 aurally distinct phrases suitable for non-visual text entry research.
  • Proposed nine transcription error categories, including comprehension errors, spelling errors, extra space errors, numeric errors, etc., and established classification standards.

• Advantages Over Existing Solutions:

  • Compared to the original phrase set, the selected phrases significantly reduced comprehension errors in speech transcription.
  • Provided a replicable method for speech generation and error classification to support future corpus expansion and applied research.

• Experimental or Evaluation Results:

  • First Experiment: 96 phrases showed no comprehension errors across multiple speech versions, with an average accuracy rate as high as 77.51%.
  • Second Experiment: 92 phrases were validated by blind or low-vision users, with accuracy rates significantly improved and error rates reduced to 23.85%.

• Limitations and Future Directions:

  • Limitations:
    • Small participant sample size may not represent broader user behavior.
    • The identified corpus is limited to English and needs to support other languages in the future.
    • Manual coding may introduce subjective bias.
  • Future Directions:
    • Develop broader, cross-linguistic aurally distinct phrase corpora.
    • Explore the impact of different speech configurations in text-to-speech technology on user comprehension.
    • Test the new phrases in specific devices and non-visual interaction scenarios, such as optimizing speech rate adjustments.