Local Word Discovery for Interactive Transcription

William Lane, Steven BirdEMNLP 2021 · DOI

Abstract

This paper introduces a "local word discovery" pipeline designed to assist in transcribing low-resource, morphologically complex languages. By integrating human expertise with a computational grammar via a finite state transducer (FST), the system bridges the gap between noisy automatic phone recognition and sparse human annotations. Evaluation on the Australian language Kunwinjku demonstrates a 17% gain in transcription density and high suggestion accuracy, facilitating rapid lexicon expansion in field settings.

Introduction

Most modern speech technology relies on massive datasets of transcribed audio to train "end-to-end" models. However, for many of the world’s 7,000 languages—particularly Indigenous languages—these datasets simply do not exist. The challenge is twofold: resource scarcity and extreme morphological complexity. Languages like Kunwinjku are polysynthetic, meaning a single "word" can function like a full sentence in English, making standard language modeling difficult.

The authors address the "Boasian trilogy" gap: while linguists often possess published grammars (the rules of the language), modern NLP often ignores these rules in favor of raw data. This research asks how existing grammars can be leveraged to accelerate the creation of new texts and lexicons in low-resource environments.

Methodology

The authors propose a task called Local Word Discovery. Instead of attempting to transcribe a whole recording autonomously, the system acts as an interactive assistant for a human transcriber.

The Local Word Discovery pipeline combining phone recognition, human input, and a morphological FST.

The pipeline consists of four primary stages:

  1. Phone Recognition: Using Allosaurus, a universal phone recognizer fine-tuned on a small amount of Kunwinjku speech, audio is converted into a stream of phonetic symbols.
  2. Sparse Transcription: A human identifies a few familiar "morphs" (word pieces) within the audio segment.
  3. Morphological Transducer (FST): This core innovation implements a published grammar of Kunwinjku as a Finite State Transducer. The FST takes the known morphs and the surrounding noisy phones to suggest linguistically valid, complex words that fit the phonetic context.
  4. Interactive Loop: The system suggests completions, which the human validates. Confirmed words are then fed back into the lexicon to improve future suggestions.
Alignment of recognized morphs (L) and noisy phones (Q) to create the combined input (I) for the FST.

Experimental Setup

The study focuses on Kunwinjku (ISO: gup), an Aboriginal Australian language. The researchers used 78 minutes of phonemically-transcribed spontaneous speech for fine-tuning and an evaluation set of 126 "breath groups" from a single speaker. To simulate an early-stage project, the experiment was seeded with only the 10 most frequent morphs.

Key Results

  • Transcription Density: A single iteration of this method resulted in a 17.3% relative increase in the number of characters transcribed.
  • Suggestion Accuracy: 75.4% of the tested audio segments received at least one correct partial or full-word suggestion.
  • Lexicon Growth: The process discovered 76 new unique lexical entries from the initial seed of 10, demonstrating rapid dictionary expansion.
  • Qualitative Impact: Field testing in the Gunbalanya community showed the tool provided a "low-friction entry point" for fluent speakers who were less confident in their written literacy.

Conclusion

In low-data environments, the linguistic "rules" of a language are often more valuable than statistical patterns. The FST acts as a filter, ensuring the machine only suggests linguistically "legal" words, reducing the cognitive load on the transcriber from "writing from scratch" to "multiple choice."

Limitations include "FST hallucination," where noisy phonetic recognition leads to an overwhelming number of incorrect but legal suggestions. Future work involves integrating this pipeline with "word spotting" to automatically find other instances of newly discovered words across larger corpora.

Summary updated

BibTeX
@inproceedings{lane2021local,
  author = {William Lane and Steven Bird},
  title = {{Local Word Discovery for Interactive Transcription}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2021},
  doi = {10.18653/v1/2021.emnlp-main.157},
}
Figure