[Paper Review] Lexical Access for Speech Understanding using Minimum Message Length Encoding

Ian B. Thomas, Ingrid Zukerman+3 more|arXiv (Cornell University)|Feb 6, 2013

Algorithms and Data Compression13 references3 citations

TL;DR

This paper proposes a Minimum Message Length (MML) encoding approach to solve the lexical access problem in speech understanding by inferring the most probable word sequence from phoneme inputs. By modeling phoneme realizations and leveraging word/POS statistics from text corpora, the method achieves accurate recognition in continuous, multi-speaker speech, with a speed-optimized heuristic using phonetically similar word equivalence classes that maintains high accuracy without significant loss.

ABSTRACT

The Lexical Access Problem consists of determining the intended sequence of words corresponding to an input sequence of phonemes (basic speech sounds) that come from a low-level phoneme recognizer. In this paper we present an information-theoretic approach based on the Minimum Message Length Criterion for solving the Lexical Access Problem. We model sentences using phoneme realizations seen in training, and word and part-of-speech information obtained from text corpora. We show results on multiple-speaker, continuous, read speech and discuss a heuristic using equivalence classes of similar sounding words which speeds up the recognition process without significant deterioration in recognition accuracy.

Motivation & Objective

• To address the lexical access problem—mapping phoneme sequences to intended word sequences in speech understanding.
• To develop an information-theoretic solution using Minimum Message Length (MML) encoding for optimal word sequence inference.
• To improve recognition efficiency through a heuristic grouping of phonetically similar words into equivalence classes.
• To evaluate the method on continuous, multi-speaker, read speech data with realistic phoneme recognition outputs.
• To balance recognition accuracy and computational speed using MML-based compression principles.

Proposed method

• The method models the joint probability of word sequences and their phonetic realizations using MML, minimizing the total message length of the hypothesis and data.
• It combines phoneme-level realizations from training data with word and part-of-speech (POS) frequency distributions from text corpora.
• The MML criterion selects the word sequence that best compresses the input phoneme sequence, favoring both data fit and model simplicity.
• A heuristic groups words with similar pronunciations into equivalence classes, reducing search space during decoding.
• The system uses a probabilistic language model derived from text corpora to guide word sequence selection.
• The decoding process applies MML to rank candidate word sequences, favoring those that minimize message length.

Experimental results

Research questions

• RQ1 Can the Minimum Message Length criterion effectively solve the lexical access problem in speech understanding?
• RQ2 How does MML-based word sequence inference compare to traditional approaches in terms of accuracy and efficiency?
• RQ3 To what extent can phonetically similar words be grouped into equivalence classes without degrading recognition accuracy?
• RQ4 How well does the MML model perform on continuous, multi-speaker, read speech with noisy phoneme inputs?
• RQ5 Can the equivalence class heuristic significantly reduce computational cost while preserving recognition quality?

Key findings

• The MML-based approach achieves high accuracy in mapping phoneme sequences to correct word sequences in continuous, multi-speaker speech.
• The use of equivalence classes for similar-sounding words reduces computational complexity with minimal impact on recognition accuracy.
• The method demonstrates robust performance on read speech data, even when phoneme recognition outputs are imperfect.
• The integration of text corpus statistics (word and POS frequencies) improves the model's ability to resolve ambiguous phoneme sequences.
• The MML framework effectively balances model complexity and data fit, leading to optimal word sequence inference.
• The heuristic approach enables faster decoding, making the system more practical for real-time applications.